Hi, I want to report that I recently implemented a support for arbitrary number of parities that could be useful also for Linux RAID and Btrfs, both currently limited to double parity. In short, to generate the parity I use a Cauchy matrix specifically built to be compatible with the existing Linux parity computation, and extensible to an arbitrary number of parities. This without limitations on the number of data disks. The Cauchy matrix for six parities is: 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01... 01 02 04 08 10 20 40 80 1d 3a 74 e8 cd 87 13 26 4c 98 2d 5a b4 75... 01 f5 d2 c4 9a 71 f1 7f fc 87 c1 c6 19 2f 40 55 3d ba 53 04 9c 61... 01 bb a6 d7 c7 07 ce 82 4a 2f a5 9b b6 60 f1 ad e7 f4 06 d2 df 2e... 01 97 7f 9c 7c 18 bd a2 58 1a da 74 70 a3 e5 47 29 07 f5 80 23 e9... 01 2b 3f cf 73 2c d6 ed cb 74 15 78 8a c1 17 c9 89 68 21 ab 76 3b... You can easily recognize the first row as RAID5 based on a simple XOR, and the second row as RAID6 based on multiplications by powers of 2. The other rows are for additional parity levels and they require multiplications by arbitrary values that can be implemented using the PSHUFB instruction. The performance of triple parity with PSHUFB is comparable at an alternate triple parity implementation with the third row of coefficients set as powers of 2^-1. This alternate implementation is likely the fastest possible for CPUs without PSHUFB or similar instruction, but it has the limitation of not supporting beyond triple parity. The Cauchy matrix is instead working for any number of parities and at the same time it''s compatible with the existing first two parity levels. As far as I know, this is a kind of new result, never appeared in this list or somewhere else. You can see more details, performance results and fast implementations for up to six parity levels at: https://sourceforge.net/p/snapraid/code/ci/master/tree/raid.c This was developed as part of my hobby project SnapRAID, downloadable with full source at: http://snapraid.sourceforge.net/ Please let me know if you are interested in a potential Linux integration. I can surely help on whatever is needed. For reference, past discussions about triple parity in the linux-raid list can be found at: http://thread.gmane.org/gmane.linux.raid/34195 http://thread.gmane.org/gmane.linux.raid/37904 Ciao, Andrea -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/18/2013 02:08 PM, Andrea Mazzoleni wrote:> Hi, > > I want to report that I recently implemented a support for > arbitrary number of parities that could be useful also for Linux > RAID and Btrfs, both currently limited to double parity. > > In short, to generate the parity I use a Cauchy matrix specifically > built to be compatible with the existing Linux parity computation, > and extensible to an arbitrary number of parities. This without > limitations on the number of data disks. > > The Cauchy matrix for six parities is: > > 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01... > 01 02 04 08 10 20 40 80 1d 3a 74 e8 cd 87 13 26 4c 98 2d 5a b4 75... > 01 f5 d2 c4 9a 71 f1 7f fc 87 c1 c6 19 2f 40 55 3d ba 53 04 9c 61... > 01 bb a6 d7 c7 07 ce 82 4a 2f a5 9b b6 60 f1 ad e7 f4 06 d2 df 2e... > 01 97 7f 9c 7c 18 bd a2 58 1a da 74 70 a3 e5 47 29 07 f5 80 23 e9... > 01 2b 3f cf 73 2c d6 ed cb 74 15 78 8a c1 17 c9 89 68 21 ab 76 3b... > > You can easily recognize the first row as RAID5 based on a simple > XOR, and the second row as RAID6 based on multiplications by powers > of 2. The other rows are for additional parity levels and they > require multiplications by arbitrary values that can be implemented > using the PSHUFB instruction. >Hello, This looks very interesting indeed. Could you perhaps describe how the Cauchy matrix is derived, and under what conditions it would become singular? -hpa -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Peter, The Cauchy matrix has the mathematical property to always have itself and all submatrices not singular. So, we are sure that we can always solve the equations to recover the data disks. Besides the mathematical proof, I''ve also inverted all the 377,342,351,231 possible submatrices for up to 6 parities and 251 data disks, and got an experimental confirmation of this. The only limit is coming from the GF(2^8). You have a maximum number of disk = 2^8 + 1 - number_of_parities. For example, with 6 parities, you can have no more of 251 data disks. Over this limit it''s not possible to build a Cauchy matrix. Note that instead with a Vandermonde matrix you don''t have the guarantee to always have all the submatrices not singular. This is the reason because using power coefficients, before or late, it happens to have unsolvable equations. You can find the code that generate the Cauchy matrix with some explanation in the comments at (see the set_cauchy() function) : http://sourceforge.net/p/snapraid/code/ci/master/tree/mktables.c Ciao, Andrea On Mon, Nov 18, 2013 at 11:12 PM, H. Peter Anvin <hpa@zytor.com> wrote:> On 11/18/2013 02:08 PM, Andrea Mazzoleni wrote: >> Hi, >> >> I want to report that I recently implemented a support for >> arbitrary number of parities that could be useful also for Linux >> RAID and Btrfs, both currently limited to double parity. >> >> In short, to generate the parity I use a Cauchy matrix specifically >> built to be compatible with the existing Linux parity computation, >> and extensible to an arbitrary number of parities. This without >> limitations on the number of data disks. >> >> The Cauchy matrix for six parities is: >> >> 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01... >> 01 02 04 08 10 20 40 80 1d 3a 74 e8 cd 87 13 26 4c 98 2d 5a b4 75... >> 01 f5 d2 c4 9a 71 f1 7f fc 87 c1 c6 19 2f 40 55 3d ba 53 04 9c 61... >> 01 bb a6 d7 c7 07 ce 82 4a 2f a5 9b b6 60 f1 ad e7 f4 06 d2 df 2e... >> 01 97 7f 9c 7c 18 bd a2 58 1a da 74 70 a3 e5 47 29 07 f5 80 23 e9... >> 01 2b 3f cf 73 2c d6 ed cb 74 15 78 8a c1 17 c9 89 68 21 ab 76 3b... >> >> You can easily recognize the first row as RAID5 based on a simple >> XOR, and the second row as RAID6 based on multiplications by powers >> of 2. The other rows are for additional parity levels and they >> require multiplications by arbitrary values that can be implemented >> using the PSHUFB instruction. >> > > Hello, > > This looks very interesting indeed. Could you perhaps describe how the > Cauchy matrix is derived, and under what conditions it would become > singular? > > -hpa > >-- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/18/2013 02:35 PM, Andrea Mazzoleni wrote:> Hi Peter, > > The Cauchy matrix has the mathematical property to always have itself > and all submatrices not singular. So, we are sure that we can always > solve the equations to recover the data disks. > > Besides the mathematical proof, I''ve also inverted all the > 377,342,351,231 possible submatrices for up to 6 parities and 251 data > disks, and got an experimental confirmation of this. >Nice.> > The only limit is coming from the GF(2^8). You have a maximum number > of disk = 2^8 + 1 - number_of_parities. For example, with 6 parities, > you can have no more of 251 data disks. Over this limit it''s not > possible to build a Cauchy matrix. >251? Not 255?> Note that instead with a Vandermonde matrix you don''t have the > guarantee to always have all the submatrices not singular. This is the > reason because using power coefficients, before or late, it happens to > have unsolvable equations. > > You can find the code that generate the Cauchy matrix with some > explanation in the comments at (see the set_cauchy() function) : > > http://sourceforge.net/p/snapraid/code/ci/master/tree/mktables.cOK, need to read up on the theoretical aspects of this, but it sounds promising. -hpa -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 19/11/13 00:25, H. Peter Anvin wrote:> On 11/18/2013 02:35 PM, Andrea Mazzoleni wrote: >> Hi Peter, >> >> The Cauchy matrix has the mathematical property to always have itself >> and all submatrices not singular. So, we are sure that we can always >> solve the equations to recover the data disks. >> >> Besides the mathematical proof, I''ve also inverted all the >> 377,342,351,231 possible submatrices for up to 6 parities and 251 data >> disks, and got an experimental confirmation of this. >> > > Nice. > >> >> The only limit is coming from the GF(2^8). You have a maximum number >> of disk = 2^8 + 1 - number_of_parities. For example, with 6 parities, >> you can have no more of 251 data disks. Over this limit it''s not >> possible to build a Cauchy matrix. >> > > 251? Not 255? > >> Note that instead with a Vandermonde matrix you don''t have the >> guarantee to always have all the submatrices not singular. This is the >> reason because using power coefficients, before or late, it happens to >> have unsolvable equations. >> >> You can find the code that generate the Cauchy matrix with some >> explanation in the comments at (see the set_cauchy() function) : >> >> http://sourceforge.net/p/snapraid/code/ci/master/tree/mktables.c > > OK, need to read up on the theoretical aspects of this, but it sounds > promising. > > -hpa >Hi all, A while back I worked through the maths for a method of extending raid to multiple parities, though I never got as far as implementing it in code (other than some simple Python test code to confirm the maths). It is also missing the maths for simplified ways to recover data. I''ve posted a couple of times with this on the linux-raid mailing list (as linked in this thread) - there has certainly been some interest, but it''s not easy to turn interest into hard work! I used an obvious expansion on the existing RAID5 and RAID6 algorithms, with parity P_n being generated from powers of 2^n. This means that the triple-parity version can be implemented by simply applying the RAID6 operations twice. For a triple parity, this works well - the matrices involved are all invertible up to 255 data disks. Beyond that, however, things drop off rapidly - quad parity implemented in the same way only supports 21 data disks, and for five parity disks you need to use 0x20 (skipping 0x10) to get even 8 data disks. This means that my method would be fine for triple parity, and would also be efficient in implementation. Beyond triple parity, the simple method has size limits for four parity and is no use on anything bigger. The Cauchy matrix method lets us go beyond that (I haven''t yet studied your code and your maths - I will do so as soon as I have the chance, but I doubt if that will be before the weekend). Would it be possible to use the simple parity system for the first three parities, and Cauchy beyond that? That would give the best of both worlds. The important thing to think about here is what would actually be useful in the real world. It is always nice to have a system that can make an array with 251 data disks and 6 parities (and I certainly think the maths involved is fun), but would anyone use such a beast? Triple parity has clear use cases. As people have moved up from raid5 to raid6, "raid7" or "raid6-3p" would be an obvious next step. I also see it as being useful for maintenance on raid6 arrays - if you want to replace disks on a raid6 array you could first add a third parity disk with an asymmetric layout, then you could replace the main disks while keeping two disk redundancy at all times. Quad parity is unlikely, I think - you would need a very wide array and unusual requirements to make quad parity a better choice than a layered system of raid10 or raid15. At most, I think it would find use as a temporary security while maintaining a triple-raid array. Remember also that such an array would be painfully slow if it ever needed to rebuild data with four missing disks - and if it is then too slow to be usable, then quad parity is not a useful solution. (Obviously anyone with /real/ experience with large arrays can give better ideas here - I like the maths of multi-parity raid, but I will not it for my small arrays.) Of course I will enjoy studying your maths here, and I''ll try to give some feedback on it. But I think for implementation purposes, the simple "powers of 4" generation of triple parity would be better than using the Cauchy matrix - it is a clear step from the existing raid6, and it can work fast on a wide variety of processors (people use ARMs and other "small" cpus on raids, not just x86 with SSE3). I believe that would mean simpler code and fewer changes, which is always popular with the kernel folk. However, if it is not possible to use Cauchy matrices to get four and more parity while keeping the same first three parities, then the balance changes and a decision needs to be made - do we (the Linux kernel developers, the btrfs developers, and the users) want a simpler system that is limited to triple parity (or quad parity with 21 + 4 disks), or do we want a more complex but more flexible system? Personally, I don''t mind either way, as long as we get a good technical solution. And I''ll do what I can to help with the maths in either case. David -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Peter, Yes, 251 disks for 6 parity. To build a NxM Cauchy matrix you need to pick N+M distinct values in the GF(2^8) and we have only 2^8 == 256 available. This means that for every row we add for an extra parity level, we have to remove one of the disk columns. Note that in true, I use an Extended Cauchy matrix that gives the first row of 1 for "free". This results in N+M <= 256+1. So, DISKS = 257 - PARITY -> 251 = 257 - 6 A brief introduction of Cauchy and Extended Cauchy matrix can be found in: Vinocha, "On Generator Cauchy Matrices of GDRS/GTRS Codes", 2012 http://www.m-hikari.com/ijcms/ijcms-2012/45-48-2012/brarIJCMS45-48-2012.pdf (just check the Introduction, the rest is not related) More details can be found in: Roth, "Introduction to Coding Theory", 2006 http://carlossicoli.free.fr/R/Roth_R.-Introduction_to_coding_theory-Cambridge_University_Press%282006%29.pdf (search for "Extended Cauchy") Ciao, Andrea On Tue, Nov 19, 2013 at 12:25 AM, H. Peter Anvin <hpa@zytor.com> wrote:> On 11/18/2013 02:35 PM, Andrea Mazzoleni wrote: >> Hi Peter, >> >> The Cauchy matrix has the mathematical property to always have itself >> and all submatrices not singular. So, we are sure that we can always >> solve the equations to recover the data disks. >> >> Besides the mathematical proof, I''ve also inverted all the >> 377,342,351,231 possible submatrices for up to 6 parities and 251 data >> disks, and got an experimental confirmation of this. >> > > Nice. > >> >> The only limit is coming from the GF(2^8). You have a maximum number >> of disk = 2^8 + 1 - number_of_parities. For example, with 6 parities, >> you can have no more of 251 data disks. Over this limit it''s not >> possible to build a Cauchy matrix. >> > > 251? Not 255? > >> Note that instead with a Vandermonde matrix you don''t have the >> guarantee to always have all the submatrices not singular. This is the >> reason because using power coefficients, before or late, it happens to >> have unsolvable equations. >> >> You can find the code that generate the Cauchy matrix with some >> explanation in the comments at (see the set_cauchy() function) : >> >> http://sourceforge.net/p/snapraid/code/ci/master/tree/mktables.c > > OK, need to read up on the theoretical aspects of this, but it sounds > promising. > > -hpa > >-- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi David, Just to say that I know your good past work, and it helped me a lot. Thanks for that! Unfortunately the Cauchy matrix is not compatible with a triple parity implementation using power coefficients. They are different and incompatible roads. I partially agree on your considerations, and in fact in my sources you can also see an alternate triple parity implementation using powers of 2^-1 == 1/2 == 0x8e, intended for CPUs not supporting PSHUFB. This is faster than using powers of 2^2 == 4, because we can divide by 2 as fast as we can multiply by 2. The choice of ZFS to use powers of 4 was likely not optimal, because to multiply by 4, it has to do two multiplications by 2. Also this method doesn''t work for quad parity, because it fails with more than 16 data disks. What I tend to do not agree, is to give too importance to low end architectures, that don''t support PSHUFB or similar instruction. Such architectures can just stay with two parity levels. Consider that to have a fast recovering (running in degraded mode) you need anyway PSHUFB to have acceptable performance. In my system I can generate triple parity at 10GB/s using SSE2, but recover only at 100MB/s without SSSE3 PSHUFB. It''s a slowdown of x100! With PSHUFB is a bit better and I can recover at 500MB/s. Note also that the ARM NEON architecture introduced the VTBL instruction, and AMD introduced VPPERM, that could be used like PSHUFB. For the complexity point of view, I don''t any see difference between the two methods. They are just two matrix with different coefficients sharing the same recovering functions. The only difference is in the optimized parity generation that uses SSSE3 instead of SSE2. Anyway, I cannot tell what is the best option for Linux RAID and Btrfs. There are for sure better qualified people in this list to say that. I can just say that systems using multiple parity levels do exist, and maybe also the Linux Kernel could benefit to have such kind of support. Here some examples: Oracle/Sun, Dell/Compellent ZFS: 3 parity drives NEC HydraStor: 3 parity drives EMC/Isilon: 4 parity drives Amplidata: 4 parity drives CleverSafe: 6 parity drives StreamScale/BigParity: 7 parity drives And Btrfs with six parities would be surely cool :) Ciao, Andrea On Tue, Nov 19, 2013 at 11:16 AM, David Brown <david.brown@hesbynett.no> wrote:> On 19/11/13 00:25, H. Peter Anvin wrote: >> On 11/18/2013 02:35 PM, Andrea Mazzoleni wrote: >>> Hi Peter, >>> >>> The Cauchy matrix has the mathematical property to always have itself >>> and all submatrices not singular. So, we are sure that we can always >>> solve the equations to recover the data disks. >>> >>> Besides the mathematical proof, I''ve also inverted all the >>> 377,342,351,231 possible submatrices for up to 6 parities and 251 data >>> disks, and got an experimental confirmation of this. >>> >> >> Nice. >> >>> >>> The only limit is coming from the GF(2^8). You have a maximum number >>> of disk = 2^8 + 1 - number_of_parities. For example, with 6 parities, >>> you can have no more of 251 data disks. Over this limit it''s not >>> possible to build a Cauchy matrix. >>> >> >> 251? Not 255? >> >>> Note that instead with a Vandermonde matrix you don''t have the >>> guarantee to always have all the submatrices not singular. This is the >>> reason because using power coefficients, before or late, it happens to >>> have unsolvable equations. >>> >>> You can find the code that generate the Cauchy matrix with some >>> explanation in the comments at (see the set_cauchy() function) : >>> >>> http://sourceforge.net/p/snapraid/code/ci/master/tree/mktables.c >> >> OK, need to read up on the theoretical aspects of this, but it sounds >> promising. >> >> -hpa >> > > Hi all, > > A while back I worked through the maths for a method of extending raid > to multiple parities, though I never got as far as implementing it in > code (other than some simple Python test code to confirm the maths). It > is also missing the maths for simplified ways to recover data. I''ve > posted a couple of times with this on the linux-raid mailing list (as > linked in this thread) - there has certainly been some interest, but > it''s not easy to turn interest into hard work! > > I used an obvious expansion on the existing RAID5 and RAID6 algorithms, > with parity P_n being generated from powers of 2^n. This means that the > triple-parity version can be implemented by simply applying the RAID6 > operations twice. For a triple parity, this works well - the matrices > involved are all invertible up to 255 data disks. Beyond that, however, > things drop off rapidly - quad parity implemented in the same way only > supports 21 data disks, and for five parity disks you need to use 0x20 > (skipping 0x10) to get even 8 data disks. > > This means that my method would be fine for triple parity, and would > also be efficient in implementation. > > Beyond triple parity, the simple method has size limits for four parity > and is no use on anything bigger. The Cauchy matrix method lets us go > beyond that (I haven''t yet studied your code and your maths - I will do > so as soon as I have the chance, but I doubt if that will be before the > weekend). > > Would it be possible to use the simple parity system for the first three > parities, and Cauchy beyond that? That would give the best of both worlds. > > > > The important thing to think about here is what would actually be useful > in the real world. It is always nice to have a system that can make an > array with 251 data disks and 6 parities (and I certainly think the > maths involved is fun), but would anyone use such a beast? > > Triple parity has clear use cases. As people have moved up from raid5 > to raid6, "raid7" or "raid6-3p" would be an obvious next step. I also > see it as being useful for maintenance on raid6 arrays - if you want to > replace disks on a raid6 array you could first add a third parity disk > with an asymmetric layout, then you could replace the main disks while > keeping two disk redundancy at all times. > > Quad parity is unlikely, I think - you would need a very wide array and > unusual requirements to make quad parity a better choice than a layered > system of raid10 or raid15. At most, I think it would find use as a > temporary security while maintaining a triple-raid array. Remember also > that such an array would be painfully slow if it ever needed to rebuild > data with four missing disks - and if it is then too slow to be usable, > then quad parity is not a useful solution. > > > (Obviously anyone with /real/ experience with large arrays can give > better ideas here - I like the maths of multi-parity raid, but I will > not it for my small arrays.) > > > > Of course I will enjoy studying your maths here, and I''ll try to give > some feedback on it. But I think for implementation purposes, the > simple "powers of 4" generation of triple parity would be better than > using the Cauchy matrix - it is a clear step from the existing raid6, > and it can work fast on a wide variety of processors (people use ARMs > and other "small" cpus on raids, not just x86 with SSE3). I believe > that would mean simpler code and fewer changes, which is always popular > with the kernel folk. > > However, if it is not possible to use Cauchy matrices to get four and > more parity while keeping the same first three parities, then the > balance changes and a decision needs to be made - do we (the Linux > kernel developers, the btrfs developers, and the users) want a simpler > system that is limited to triple parity (or quad parity with 21 + 4 > disks), or do we want a more complex but more flexible system? > > Personally, I don''t mind either way, as long as we get a good technical > solution. And I''ll do what I can to help with the maths in either case. > > David > >-- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Mon, Nov 18, 2013 at 11:08:59PM +0100, Andrea Mazzoleni wrote:> Hi, > > I want to report that I recently implemented a support for > arbitrary number of parities that could be useful also for Linux > RAID and Btrfs, both currently limited to double parity. > > In short, to generate the parity I use a Cauchy matrix specifically > built to be compatible with the existing Linux parity computation, > and extensible to an arbitrary number of parities. This without > limitations on the number of data disks. > > The Cauchy matrix for six parities is: > > 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01 01... > 01 02 04 08 10 20 40 80 1d 3a 74 e8 cd 87 13 26 4c 98 2d 5a b4 75... > 01 f5 d2 c4 9a 71 f1 7f fc 87 c1 c6 19 2f 40 55 3d ba 53 04 9c 61... > 01 bb a6 d7 c7 07 ce 82 4a 2f a5 9b b6 60 f1 ad e7 f4 06 d2 df 2e... > 01 97 7f 9c 7c 18 bd a2 58 1a da 74 70 a3 e5 47 29 07 f5 80 23 e9... > 01 2b 3f cf 73 2c d6 ed cb 74 15 78 8a c1 17 c9 89 68 21 ab 76 3b... > > You can easily recognize the first row as RAID5 based on a simple > XOR, and the second row as RAID6 based on multiplications by powers > of 2. The other rows are for additional parity levels and they > require multiplications by arbitrary values that can be implemented > using the PSHUFB instruction. > > The performance of triple parity with PSHUFB is comparable at an > alternate triple parity implementation with the third row of > coefficients set as powers of 2^-1. This alternate implementation is > likely the fastest possible for CPUs without PSHUFB or similar > instruction, but it has the limitation of not supporting beyond triple > parity. > > The Cauchy matrix is instead working for any number of parities and > at the same time it''s compatible with the existing first two parity > levels. As far as I know, this is a kind of new result, never > appeared in this list or somewhere else. > > You can see more details, performance results and fast > implementations for up to six parity levels at: > https://sourceforge.net/p/snapraid/code/ci/master/tree/raid.c > > This was developed as part of my hobby project SnapRAID, > downloadable with full source at: > http://snapraid.sourceforge.net/ > > Please let me know if you are interested in a potential Linux > integration. I can surely help on whatever is needed. > > For reference, past discussions about triple parity in the > linux-raid list can be found at: > > http://thread.gmane.org/gmane.linux.raid/34195 > http://thread.gmane.org/gmane.linux.raid/37904Hi Andrea, great job, this was exactly what I was looking for. Do you know if there is a "fast" way not to correct errors, but to find them? In RAID-6 (as per raid6check) there is an easy way to verify where an HDD has incorrect data. I suspect, for each 2 parity block it should be possible to find 1 error (and if this is true, then quad parity is more attractive than triple one). Furthermore, my second (of first) target would be something like: http://www.symform.com/blog/tag/raid-96/ Which uses 32 parities (out of 96 "disks"). Keep going!!! bye, pg> > Ciao, > Andrea > -- > To unsubscribe from this list: send the line "unsubscribe linux-raid" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html-- piergiorgio -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/19/2013 12:28 PM, Andrea Mazzoleni wrote:> Hi Peter, > > Yes, 251 disks for 6 parity. > > To build a NxM Cauchy matrix you need to pick N+M distinct values > in the GF(2^8) and we have only 2^8 == 256 available. > This means that for every row we add for an extra parity level, we > have to remove one of the disk columns. > > Note that in true, I use an Extended Cauchy matrix that gives the > first row of 1 for "free". This results in N+M <= 256+1. > So, DISKS = 257 - PARITY -> 251 = 257 - 6 > > A brief introduction of Cauchy and Extended Cauchy matrix can be found in: > > Vinocha, "On Generator Cauchy Matrices of GDRS/GTRS Codes", 2012 > http://www.m-hikari.com/ijcms/ijcms-2012/45-48-2012/brarIJCMS45-48-2012.pdf > (just check the Introduction, the rest is not related) > > More details can be found in: > > Roth, "Introduction to Coding Theory", 2006 > http://carlossicoli.free.fr/R/Roth_R.-Introduction_to_coding_theory-Cambridge_University_Press%282006%29.pdf > (search for "Extended Cauchy") > > Ciao, > AndreaGreat work - we have waited a long time for this. Adding in Jim Plank who did some great talks and work in this area as well :) Ric> > On Tue, Nov 19, 2013 at 12:25 AM, H. Peter Anvin <hpa@zytor.com> wrote: >> On 11/18/2013 02:35 PM, Andrea Mazzoleni wrote: >>> Hi Peter, >>> >>> The Cauchy matrix has the mathematical property to always have itself >>> and all submatrices not singular. So, we are sure that we can always >>> solve the equations to recover the data disks. >>> >>> Besides the mathematical proof, I''ve also inverted all the >>> 377,342,351,231 possible submatrices for up to 6 parities and 251 data >>> disks, and got an experimental confirmation of this. >>> >> Nice. >> >>> The only limit is coming from the GF(2^8). You have a maximum number >>> of disk = 2^8 + 1 - number_of_parities. For example, with 6 parities, >>> you can have no more of 251 data disks. Over this limit it''s not >>> possible to build a Cauchy matrix. >>> >> 251? Not 255? >> >>> Note that instead with a Vandermonde matrix you don''t have the >>> guarantee to always have all the submatrices not singular. This is the >>> reason because using power coefficients, before or late, it happens to >>> have unsolvable equations. >>> >>> You can find the code that generate the Cauchy matrix with some >>> explanation in the comments at (see the set_cauchy() function) : >>> >>> http://sourceforge.net/p/snapraid/code/ci/master/tree/mktables.c >> OK, need to read up on the theoretical aspects of this, but it sounds >> promising. >> >> -hpa >> >> > -- > To unsubscribe from this list: send the line "unsubscribe linux-raid" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html-- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
I''m not going to claim any expert status on this discussion (the theory makes my head spin) but I will say I agree with Andrea as far as prefering his implementation for triple parity and beyond. PSHUFB has been around the intel platform since the Core2 introduced it as part of SSSE3 back in Q1 2006. The generation of Intel based servers that ran pre-Core Xeons are long in the tooth and this is a value judgement but if your data is big enough you need triple parity, you probably shouldn''t be running it from an ten year old platform. But that''s just me. :-) -- Drew "Nothing in life is to be feared. It is only to be understood." --Marie Curie "This started out as a hobby and spun horribly out of control." -Unknown -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Nov 19, 2013, at 3:51 PM, Drew <drew.kay@gmail.com> wrote:> I''m not going to claim any expert status on this discussion (the > theory makes my head spin) but I will say I agree with Andrea as far > as prefering his implementation for triple parity and beyond. > > PSHUFB has been around the intel platform since the Core2 introduced > it as part of SSSE3 back in Q1 2006. The generation of Intel based > servers that ran pre-Core Xeons are long in the tooth and this is a > value judgement but if your data is big enough you need triple parity, > you probably shouldn''t be running it from an ten year old platform.If anything, I''d like to see two implementations of RAID 6 dual parity. The existing implementation in the md driver and btrfs could remain the default, but users could opt into Cauchy matrix based dual parity which would then enable them an easy (and live) migration path to triple parity and beyond. Chris Murphy-- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, Nov 19, 2013 at 4:54 PM, Chris Murphy <lists@colorremedies.com> wrote:> If anything, I''d like to see two implementations of RAID 6 dual > parity. The existing implementation in the md driver and btrfs could > remain the default, but users could opt into Cauchy matrix based dual > parity which would then enable them an easy (and live) migration path > to triple parity and beyond.Actually, my understanding is that Andrea''s Cauchy matrix technique (call it C) is compatible with existing md RAID5 and RAID6 (call these A). It is only the non-SSSE3 triple-parity algorithm 2^-1 (call it B) that is incompatible with his Cauchy matrix technique. So, you can have: 1) A+B or 2) A+C But you cannot have A+B+C -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 19/11/13 18:36, Andrea Mazzoleni wrote:> Hi David, > > Just to say that I know your good past work, and it helped me a lot. > Thanks for that!If we end up with your Cauchy matrix implementation going into the kernel and btrfs (and you''ve persuaded me, anyway), then perhaps I should make a new version of that document with the Cauchy matrix maths. I know I have found Peter''s original Raid 6 document very useful and informative, and I think it would be good to have an equivalent here too. A lot of people find the maths of Raid 6 hard to grasp - it doesn''t get easier with triple parity, and such documentation could help future kernel/btrfs developers.> > Unfortunately the Cauchy matrix is not compatible with a triple parity > implementation using power coefficients. They are different and > incompatible roads.Yes. I haven''t worked through all the details of how Cauchy matrices work - but I have read enough to see that this is correct. I am sure that it is /possible/ to make matrices that have the first three rows matching the power coefficient triple parity with at least 6 rows in total and a useful horizontal size that have the required property (all square submatrices are invertible) - but I have no idea how to find such matrices in a sensible time frame. The key point with the Cauchy matrix is not the particular form of the coefficients, but that it gives a way to generate coefficients easily.> > I partially agree on your considerations, and in fact in my sources > you can also see an alternate triple parity implementation using powers > of 2^-1 == 1/2 == 0x8e, intended for CPUs not supporting PSHUFB. > This is faster than using powers of 2^2 == 4, because we can divide > by 2 as fast as we can multiply by 2. > The choice of ZFS to use powers of 4 was likely not optimal, > because to multiply by 4, it has to do two multiplications by 2.I can agree with that. I didn''t copy ZFS''s choice here (I knew that ZFS /had/ triple parity, but I think my suggestion is not exactly the same) - I just reached the same conclusion that since we have an operation that works fast, doing it twice is not going to be an unreasonable cost. It hadn''t occurred to me that dividing by 2 was equally easy.> Also this method doesn''t work for quad parity, because it fails with > more than 16 data disks.Indeed.> > What I tend to do not agree, is to give too importance to low end > architectures, that don''t support PSHUFB or similar instruction. > Such architectures can just stay with two parity levels.That''s certainly a reasonable way to look at it. We should not limit the possibilities for high-end systems because of the limitations of low-end systems that are unlikely to use 3+ parity anyway. I''ve also looked up a list of the processors that support SSE3 and PSHUFB - a lot of modern "low-end" x86 cpus support it. And of course it is possible to implement general G(2^8) multiplication without PSHUFB, using a lookup table - it is important that this can all work with any CPU, even if it is slow.> Consider that to have a fast recovering (running in degraded mode) > you need anyway PSHUFB to have acceptable performance.I don''t think we need to be very concerned with fast recovery using the third (or more) parity. If one disk (or sector on a disk) has failed, recovery is done with the RAID5 parity block. If we have two failures, we can use the RAID6 procedure (I haven''t looked at the implementation for this at the moment, or for your PSHUFB code - but since RAID6 recovery also involves general multiplication, there may be scope for speedups here). It is only with three simultaneous failures that we need to use additional parities - such situations should be very rare. Typically this would be only for stripes with unrecoverable read errors found while rebuilding for another missed disk or two. Of course, it is always nice to have fast recovery - but fast generation of parities is far more important.> In my system I can generate triple parity at 10GB/s using SSE2, > but recover only at 100MB/s without SSSE3 PSHUFB. It''s a slowdown > of x100! With PSHUFB is a bit better and I can recover at 500MB/s.The difference here is bigger than I would have guessed, but I haven''t looked at the code yet.> Note also that the ARM NEON architecture introduced the VTBL > instruction, and AMD introduced VPPERM, that could be used like > PSHUFB.Sounds good - at least it will be fun trying to figure out optimal code.> > For the complexity point of view, I don''t any see difference between > the two methods. > They are just two matrix with different coefficients sharing the same > recovering functions. The only difference is in the optimized parity > generation that uses SSSE3 instead of SSE2.Recovery is the same, yes. It is only the parity generation that is different - multiplying by powers of 2 means each step is a fast multiply-by-two, with Horner''s rule to avoid any other multiplication. With parity 3 generated as powers of 4 or 2^-1, you have the same system with only a slightly slower multiply-by-4 step. With the Cauchy matrix, you need general multiplication with different coefficients for each disk block. This is significantly more complex - but if it can be done fast enough on at least a reasonable selection of processors, it''s okay to be complex.> > Anyway, I cannot tell what is the best option for Linux RAID and Btrfs. > There are for sure better qualified people in this list to say that.I think H. Peter Anvin is the best qualified for such decisions - I believe he has the most experience and understanding in this area. For what it is worth, you have convinced /me/ that your Cauchy matrices are the way to go. I will want to study your code a bit more, and try it out for myself, but it looks like you have a way to overcome the limitations of the power sequence method without too big runtime costs - and that is exactly what we need.> I can just say that systems using multiple parity levels do exist, and > maybe also the Linux Kernel could benefit to have such kind of support.I certainly think so. I think 3 parities is definitely useful, and sometimes four would be nice. Beyond that, I suspect "coolness" and bragging rights (btrfs can support more parities than ZFS...) will outweigh real-life implementations, so it is important that the implementation does not sacrifice anything on the triple parity in order to get 5+ parity support. It''s fine for /us/ to look at fun solutions, but it needs to be practical too if it is going to be accepted in the kernel. mvh., David> > Here some examples: > > Oracle/Sun, Dell/Compellent ZFS: 3 parity drives > NEC HydraStor: 3 parity drives > EMC/Isilon: 4 parity drives > Amplidata: 4 parity drives > CleverSafe: 6 parity drives > StreamScale/BigParity: 7 parity drives > > And Btrfs with six parities would be surely cool :) > > Ciao, > Andrea > > On Tue, Nov 19, 2013 at 11:16 AM, David Brown <david.brown@hesbynett.no> wrote: >> On 19/11/13 00:25, H. Peter Anvin wrote: >>> On 11/18/2013 02:35 PM, Andrea Mazzoleni wrote: >>>> Hi Peter, >>>> >>>> The Cauchy matrix has the mathematical property to always have itself >>>> and all submatrices not singular. So, we are sure that we can always >>>> solve the equations to recover the data disks. >>>> >>>> Besides the mathematical proof, I''ve also inverted all the >>>> 377,342,351,231 possible submatrices for up to 6 parities and 251 data >>>> disks, and got an experimental confirmation of this. >>>> >>> >>> Nice. >>> >>>> >>>> The only limit is coming from the GF(2^8). You have a maximum number >>>> of disk = 2^8 + 1 - number_of_parities. For example, with 6 parities, >>>> you can have no more of 251 data disks. Over this limit it''s not >>>> possible to build a Cauchy matrix. >>>> >>> >>> 251? Not 255? >>> >>>> Note that instead with a Vandermonde matrix you don''t have the >>>> guarantee to always have all the submatrices not singular. This is the >>>> reason because using power coefficients, before or late, it happens to >>>> have unsolvable equations. >>>> >>>> You can find the code that generate the Cauchy matrix with some >>>> explanation in the comments at (see the set_cauchy() function) : >>>> >>>> http://sourceforge.net/p/snapraid/code/ci/master/tree/mktables.c >>> >>> OK, need to read up on the theoretical aspects of this, but it sounds >>> promising. >>> >>> -hpa >>> >> >> Hi all, >> >> A while back I worked through the maths for a method of extending raid >> to multiple parities, though I never got as far as implementing it in >> code (other than some simple Python test code to confirm the maths). It >> is also missing the maths for simplified ways to recover data. I''ve >> posted a couple of times with this on the linux-raid mailing list (as >> linked in this thread) - there has certainly been some interest, but >> it''s not easy to turn interest into hard work! >> >> I used an obvious expansion on the existing RAID5 and RAID6 algorithms, >> with parity P_n being generated from powers of 2^n. This means that the >> triple-parity version can be implemented by simply applying the RAID6 >> operations twice. For a triple parity, this works well - the matrices >> involved are all invertible up to 255 data disks. Beyond that, however, >> things drop off rapidly - quad parity implemented in the same way only >> supports 21 data disks, and for five parity disks you need to use 0x20 >> (skipping 0x10) to get even 8 data disks. >> >> This means that my method would be fine for triple parity, and would >> also be efficient in implementation. >> >> Beyond triple parity, the simple method has size limits for four parity >> and is no use on anything bigger. The Cauchy matrix method lets us go >> beyond that (I haven''t yet studied your code and your maths - I will do >> so as soon as I have the chance, but I doubt if that will be before the >> weekend). >> >> Would it be possible to use the simple parity system for the first three >> parities, and Cauchy beyond that? That would give the best of both worlds. >> >> >> >> The important thing to think about here is what would actually be useful >> in the real world. It is always nice to have a system that can make an >> array with 251 data disks and 6 parities (and I certainly think the >> maths involved is fun), but would anyone use such a beast? >> >> Triple parity has clear use cases. As people have moved up from raid5 >> to raid6, "raid7" or "raid6-3p" would be an obvious next step. I also >> see it as being useful for maintenance on raid6 arrays - if you want to >> replace disks on a raid6 array you could first add a third parity disk >> with an asymmetric layout, then you could replace the main disks while >> keeping two disk redundancy at all times. >> >> Quad parity is unlikely, I think - you would need a very wide array and >> unusual requirements to make quad parity a better choice than a layered >> system of raid10 or raid15. At most, I think it would find use as a >> temporary security while maintaining a triple-raid array. Remember also >> that such an array would be painfully slow if it ever needed to rebuild >> data with four missing disks - and if it is then too slow to be usable, >> then quad parity is not a useful solution. >> >> >> (Obviously anyone with /real/ experience with large arrays can give >> better ideas here - I like the maths of multi-parity raid, but I will >> not it for my small arrays.) >> >> >> >> Of course I will enjoy studying your maths here, and I''ll try to give >> some feedback on it. But I think for implementation purposes, the >> simple "powers of 4" generation of triple parity would be better than >> using the Cauchy matrix - it is a clear step from the existing raid6, >> and it can work fast on a wide variety of processors (people use ARMs >> and other "small" cpus on raids, not just x86 with SSE3). I believe >> that would mean simpler code and fewer changes, which is always popular >> with the kernel folk. >> >> However, if it is not possible to use Cauchy matrices to get four and >> more parity while keeping the same first three parities, then the >> balance changes and a decision needs to be made - do we (the Linux >> kernel developers, the btrfs developers, and the users) want a simpler >> system that is limited to triple parity (or quad parity with 21 + 4 >> disks), or do we want a more complex but more flexible system? >> >> Personally, I don''t mind either way, as long as we get a good technical >> solution. And I''ll do what I can to help with the maths in either case. >> >> David >> >> >-- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 20/11/13 02:23, John Williams wrote:> On Tue, Nov 19, 2013 at 4:54 PM, Chris Murphy <lists@colorremedies.com> > wrote: >> If anything, I''d like to see two implementations of RAID 6 dual >> parity. The existing implementation in the md driver and btrfs could >> remain the default, but users could opt into Cauchy matrix based dual >> parity which would then enable them an easy (and live) migration path >> to triple parity and beyond.Andrea''s Cauchy matrix is compatible with the existing Raid6, so there is no problem there. I believe it would be a terrible idea to have an incompatible extension - that would mean you could not have temporary extra parity drives with asymmetrical layouts, which is something I see as a very useful feature.> > Actually, my understanding is that Andrea''s Cauchy matrix technique > (call it C) is compatible with existing md RAID5 and RAID6 (call these > A). It is only the non-SSSE3 triple-parity algorithm 2^-1 (call it B) > that is incompatible with his Cauchy matrix technique. > > So, you can have: > > 1) A+B > > or > > 2) A+C > > But you cannot have A+B+CYes, that''s right. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 19/11/13 19:12, Piergiorgio Sartor wrote:> On Mon, Nov 18, 2013 at 11:08:59PM +0100, Andrea Mazzoleni wrote:<snip for brevity>> > Hi Andrea, > > great job, this was exactly what I was looking for. > > Do you know if there is a "fast" way not to correct > errors, but to find them? > > In RAID-6 (as per raid6check) there is an easy way > to verify where an HDD has incorrect data. >I think the way to do that is just to generate the parity blocks from the data blocks, and compare them to the existing parity blocks.> I suspect, for each 2 parity block it should be > possible to find 1 error (and if this is true, then > quad parity is more attractive than triple one). > > Furthermore, my second (of first) target would > be something like: http://www.symform.com/blog/tag/raid-96/ > > Which uses 32 parities (out of 96 "disks").I believe Andrea''s matrix is extensible as long as you have no more than 257 disks in total. A mere 32 parities should not be a problem :-) mvh., David> > Keep going!!! > > bye, > > pg > >> >> Ciao, >> Andrea >> -- >> To unsubscribe from this list: send the line "unsubscribe linux-raid" in >> the body of a message to majordomo@vger.kernel.org >> More majordomo info at http://vger.kernel.org/majordomo-info.html >-- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi all -- no real comments, except as I mentioned to Ric, my tutorial in FAST last February presents Reed-Solomon coding with Cauchy matrices, and then makes special note of the common pitfall of assuming that you can append a Vandermonde matrix to an identity matrix. Please see http://web.eecs.utk.edu/~plank/plank/papers/2013-02-11-FAST-Tutorial.pdf, slides 48-52. Andrea, does the matrix that you included in an earlier mail (the one that has Linux RAID-6 in the first two rows) have a general form, or did you develop it in an ad hoc manner so that it would include Linux RAID-6 in the first two rows? Best wishes -- Jim ---------- On Nov 19, 2013, at 3:29 PM, Ric Wheeler wrote:> On 11/19/2013 12:28 PM, Andrea Mazzoleni wrote: >> Hi Peter, >> >> Yes, 251 disks for 6 parity. >> >> To build a NxM Cauchy matrix you need to pick N+M distinct values >> in the GF(2^8) and we have only 2^8 == 256 available. >> This means that for every row we add for an extra parity level, we >> have to remove one of the disk columns. >> >> Note that in true, I use an Extended Cauchy matrix that gives the >> first row of 1 for "free". This results in N+M <= 256+1. >> So, DISKS = 257 - PARITY -> 251 = 257 - 6 >> >> A brief introduction of Cauchy and Extended Cauchy matrix can be found in: >> >> Vinocha, "On Generator Cauchy Matrices of GDRS/GTRS Codes", 2012 >> http://www.m-hikari.com/ijcms/ijcms-2012/45-48-2012/brarIJCMS45-48-2012.pdf >> (just check the Introduction, the rest is not related) >> >> More details can be found in: >> >> Roth, "Introduction to Coding Theory", 2006 >> http://carlossicoli.free.fr/R/Roth_R.-Introduction_to_coding_theory-Cambridge_University_Press%282006%29.pdf >> (search for "Extended Cauchy") >> >> Ciao, >> Andrea > > Great work - we have waited a long time for this. Adding in Jim Plank who did some great talks and work in this area as well :) > > Ric > >> >> On Tue, Nov 19, 2013 at 12:25 AM, H. Peter Anvin <hpa@zytor.com> wrote: >>> On 11/18/2013 02:35 PM, Andrea Mazzoleni wrote: >>>> Hi Peter, >>>> >>>> The Cauchy matrix has the mathematical property to always have itself >>>> and all submatrices not singular. So, we are sure that we can always >>>> solve the equations to recover the data disks. >>>> >>>> Besides the mathematical proof, I''ve also inverted all the >>>> 377,342,351,231 possible submatrices for up to 6 parities and 251 data >>>> disks, and got an experimental confirmation of this. >>>> >>> Nice. >>> >>>> The only limit is coming from the GF(2^8). You have a maximum number >>>> of disk = 2^8 + 1 - number_of_parities. For example, with 6 parities, >>>> you can have no more of 251 data disks. Over this limit it''s not >>>> possible to build a Cauchy matrix. >>>> >>> 251? Not 255? >>> >>>> Note that instead with a Vandermonde matrix you don''t have the >>>> guarantee to always have all the submatrices not singular. This is the >>>> reason because using power coefficients, before or late, it happens to >>>> have unsolvable equations. >>>> >>>> You can find the code that generate the Cauchy matrix with some >>>> explanation in the comments at (see the set_cauchy() function) : >>>> >>>> http://sourceforge.net/p/snapraid/code/ci/master/tree/mktables.c >>> OK, need to read up on the theoretical aspects of this, but it sounds >>> promising. >>> >>> -hpa >>> >>> >> -- >> To unsubscribe from this list: send the line "unsubscribe linux-raid" in >> the body of a message to majordomo@vger.kernel.org >> More majordomo info at http://vger.kernel.org/majordomo-info.html >-- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Nov 20, 2013 at 2:31 AM, David Brown <david.brown@hesbynett.no> wrote:> That''s certainly a reasonable way to look at it. We should not limit > the possibilities for high-end systems because of the limitations of > low-end systems that are unlikely to use 3+ parity anyway. I''ve also > looked up a list of the processors that support SSE3 and PSHUFB - a lot > of modern "low-end" x86 cpus support it. And of course it is possible > to implement general G(2^8) multiplication without PSHUFB, using a > lookup table - it is important that this can all work with any CPU, even > if it is slow.Unfortunately, it is SSSE3 that is required for PSHUFB. The SSE3 set with only two-esses does not suffice. I made that same mistake when I first heard about Andrea''s 6-parity work. SSSE3 vs. SSE3, confusing notation! SSSE3 is significantly less widely supported than SSE3. Particularly on AMD, only the very latest CPUs seem to support SSSE3. Intel support for SSSE3 goes back much further than AMD support. Maybe it is not such a big problem, since it may be possible to support two "roads". Both roads would include the current md RAID-5 and RAID-6. But one road, which those lacking CPUs supporting SSSE3 might choose, would continue on to the non-SSSE3 triple-parity 2^-1 technique, and then dead-end. The other road would continue with the Cauchy matrix technique through 3-parity all the way to 6-parity. It might even be feasible to allow someone stuck at the end of the non-SSSE3 road to convert to the Cauchy road. You would have to go through all the 2^-1 triple-parity and convert it to Cauchy triple-parity. But then you would be safely on the Cauchy road. -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi David,>> The choice of ZFS to use powers of 4 was likely not optimal, >> because to multiply by 4, it has to do two multiplications by 2. > I can agree with that. I didn''t copy ZFS''s choice hereDavid, it was not my intention to suggest that you copied from ZFS. Sorry to have expressed myself badly. I just mentioned ZFS because it''s an implementation that I know uses powers of 4 to generate triple parity, and I saw in the code that it''s implemented with two multiplication by 2. Ciao, Andrea -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
It is also possible to quickly multiply by 2^-1 which makes for an interesting R parity. Andrea Mazzoleni <amadvance@gmail.com> wrote:>Hi David, > >>> The choice of ZFS to use powers of 4 was likely not optimal, >>> because to multiply by 4, it has to do two multiplications by 2. >> I can agree with that. I didn''t copy ZFS''s choice here >David, it was not my intention to suggest that you copied from ZFS. >Sorry to have expressed myself badly. I just mentioned ZFS because it''s >an implementation that I know uses powers of 4 to generate triple >parity, and I saw in the code that it''s implemented with two >multiplication >by 2. > >Ciao, >Andrea-- Sent from my mobile phone. Please pardon brevity and lack of formatting. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi John, Yes. There are still AMD CPUs sold without SSSE3. Most notably Athlon. Instead, Intel is providing SSSE3 from the Core 2 Duo. A detailed list is available at: http://en.wikipedia.org/wiki/SSSE3 Ciao, Andrea On Wed, Nov 20, 2013 at 7:09 PM, John Williams <jwilliams4200@gmail.com> wrote:> On Wed, Nov 20, 2013 at 2:31 AM, David Brown <david.brown@hesbynett.no> wrote: >> That''s certainly a reasonable way to look at it. We should not limit >> the possibilities for high-end systems because of the limitations of >> low-end systems that are unlikely to use 3+ parity anyway. I''ve also >> looked up a list of the processors that support SSE3 and PSHUFB - a lot >> of modern "low-end" x86 cpus support it. And of course it is possible >> to implement general G(2^8) multiplication without PSHUFB, using a >> lookup table - it is important that this can all work with any CPU, even >> if it is slow. > > Unfortunately, it is SSSE3 that is required for PSHUFB. The SSE3 set > with only two-esses does not suffice. I made that same mistake when I > first heard about Andrea''s 6-parity work. SSSE3 vs. SSE3, confusing > notation! > > SSSE3 is significantly less widely supported than SSE3. Particularly > on AMD, only the very latest CPUs seem to support SSSE3. Intel support > for SSSE3 goes back much further than AMD support. > > Maybe it is not such a big problem, since it may be possible to > support two "roads". Both roads would include the current md RAID-5 > and RAID-6. But one road, which those lacking CPUs supporting SSSE3 > might choose, would continue on to the non-SSSE3 triple-parity 2^-1 > technique, and then dead-end. The other road would continue with the > Cauchy matrix technique through 3-parity all the way to 6-parity. > > It might even be feasible to allow someone stuck at the end of the > non-SSSE3 road to convert to the Cauchy road. You would have to go > through all the 2^-1 triple-parity and convert it to Cauchy > triple-parity. But then you would be safely on the Cauchy road.-- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi, Yep. At present to multiply for 2^-1 I''m using in C: static inline uint64_t d2_64(uint64_t v) { uint64_t mask = v & 0x0101010101010101U; mask = (mask << 8) - mask; v = (v >> 1) & 0x7f7f7f7f7f7f7f7fU; v ^= mask & 0x8e8e8e8e8e8e8e8eU; return v; } and for SSE2: asm volatile("movdqa %xmm2,%xmm4"); asm volatile("pxor %xmm5,%xmm5"); asm volatile("psllw $7,%xmm4"); asm volatile("psrlw $1,%xmm2"); asm volatile("pcmpgtb %xmm4,%xmm5"); asm volatile("pand %xmm6,%xmm2"); with xmm6 == 7f7f7f7f7f7f... asm volatile("pand %xmm3,%xmm5"); with xmm3 == 8e8e8e8e8e... asm volatile("pxor %xmm5,%xmm2"); where xmm2 is the intput/output Ciao, Andrea On Wed, Nov 20, 2013 at 7:43 PM, H. Peter Anvin <hpa@zytor.com> wrote:> It is also possible to quickly multiply by 2^-1 which makes for an interesting R parity. > > Andrea Mazzoleni <amadvance@gmail.com> wrote: >>Hi David, >> >>>> The choice of ZFS to use powers of 4 was likely not optimal, >>>> because to multiply by 4, it has to do two multiplications by 2. >>> I can agree with that. I didn''t copy ZFS''s choice here >>David, it was not my intention to suggest that you copied from ZFS. >>Sorry to have expressed myself badly. I just mentioned ZFS because it''s >>an implementation that I know uses powers of 4 to generate triple >>parity, and I saw in the code that it''s implemented with two >>multiplication >>by 2. >> >>Ciao, >>Andrea > > -- > Sent from my mobile phone. Please pardon brevity and lack of formatting.-- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/20/2013 10:56 AM, Andrea Mazzoleni wrote:> Hi, > > Yep. At present to multiply for 2^-1 I''m using in C: > > static inline uint64_t d2_64(uint64_t v) > { > uint64_t mask = v & 0x0101010101010101U; > mask = (mask << 8) - mask; > v = (v >> 1) & 0x7f7f7f7f7f7f7f7fU; > v ^= mask & 0x8e8e8e8e8e8e8e8eU; > return v; > } > > and for SSE2: > > asm volatile("movdqa %xmm2,%xmm4"); > asm volatile("pxor %xmm5,%xmm5"); > asm volatile("psllw $7,%xmm4"); > asm volatile("psrlw $1,%xmm2"); > asm volatile("pcmpgtb %xmm4,%xmm5"); > asm volatile("pand %xmm6,%xmm2"); with xmm6 == 7f7f7f7f7f7f... > asm volatile("pand %xmm3,%xmm5"); with xmm3 == 8e8e8e8e8e... > asm volatile("pxor %xmm5,%xmm2"); > > where xmm2 is the intput/output >Now, that doesn''t sound like something that can get neatly meshed into the Cauchy matrix scheme, I assume. It is somewhat nice to have a scheme which is arbitrarily expandable without having to fall back to dual parity during the restripe operation. It probably also reduces the amount of code necessary. -hpa -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/20/2013 10:56 AM, Andrea Mazzoleni wrote:> Hi, > > Yep. At present to multiply for 2^-1 I''m using in C: > > static inline uint64_t d2_64(uint64_t v) > { > uint64_t mask = v & 0x0101010101010101U; > mask = (mask << 8) - mask;(mask << 7) I assume... -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Jim, I build the matrix in a way that results in coefficients matching Linux RAID for the first two rows, and at the same time gives the guarantee that all the square submatrices are not singular, resulting in a MDS code. I start forming a Cauchy matrix setting each element to 1/(xi+yj) where all xi and yj are distinct elements. This is how a Cauchy matrix is usually defined in textbooks. For the first row with j=0, I use xi = 2^-i and y0 = 0, that results in: row j=0 -> 1/(xi+y0) = 1/(2^-i + 0) = 2^i (RAID-6 coefficients) For the next rows with j>0, I use yj = 2^j, resulting in: rows j>0 -> 1/(xi+yj) = 1/(2^-i + 2^j) with xi != yj for any i,j with i>=0,j>=1,i+j<255 Then I put at the top of the Cauchy matrix a row filled with 1, transforming it in an Extended Cauchy Matrix. This transformation maintains the property of having all the square submatrices not singular. I found this property mentioned in some papers/textbooks, like in the introduction of: Vinocha, On Generator Cauchy Matrices of GDRS/GTRS Codes, 2012 http://www.m-hikari.com/ijcms/ijcms-2012/45-48-2012/brarIJCMS45-48-2012.pdf Finally I adjust all the rows to have the first column filled with 1, with a multiplication of each row for an adjusting factor. Also this transformation maintains the property of having all the square submatrices not singular, and then we have a MDS code. Ciao, Andrea -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/20/2013 11:05 AM, Andrea Mazzoleni wrote:> > For the first row with j=0, I use xi = 2^-i and y0 = 0, that results in: >How can xi = 2^-i if x is supposed to be constant? That doesn''t mean that your approach isn''t valid, of course, but it might not be a Cauchy matrix and thus needs additional analysis.> row j=0 -> 1/(xi+y0) = 1/(2^-i + 0) = 2^i (RAID-6 coefficients) > > For the next rows with j>0, I use yj = 2^j, resulting in: > > rows j>0 -> 1/(xi+yj) = 1/(2^-i + 2^j)Even more so here... 2^-i and 2^j don''t seem to be of the form xi and yj respectively. -hpa -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Peter, I think I understand it differently. Concrete example in GF(256) for k=6, m=4: First, create a 3 by 6 cauchy matrix, using x_i = 2^-i, and y_i = 0 for i=0, and y_i = 2^i for other i. In this case: x = { 1, 142, 71, 173, 216, 108 } y = { 0, 2, 4). The cauchy matrix is: 1 2 4 8 16 32 244 83 78 183 118 47 167 39 213 59 153 82 Divide row 2 by 244 and row 3 by 167. Then extend it with a row of ones on top and it''s still MDS, and that''s the code for m=4, with RAID-6 as a subset. Very nice! Jim ---------- On Nov 20, 2013, at 2:10 PM, H. Peter Anvin wrote:> On 11/20/2013 11:05 AM, Andrea Mazzoleni wrote: >> >> For the first row with j=0, I use xi = 2^-i and y0 = 0, that results in: >> > > How can xi = 2^-i if x is supposed to be constant? > > That doesn''t mean that your approach isn''t valid, of course, but it > might not be a Cauchy matrix and thus needs additional analysis. > >> row j=0 -> 1/(xi+y0) = 1/(2^-i + 0) = 2^i (RAID-6 coefficients) >> >> For the next rows with j>0, I use yj = 2^j, resulting in: >> >> rows j>0 -> 1/(xi+yj) = 1/(2^-i + 2^j) > > Even more so here... 2^-i and 2^j don''t seem to be of the form xi and yj > respectively. > > -hpa >-- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Peter,>> static inline uint64_t d2_64(uint64_t v) >> { >> uint64_t mask = v & 0x0101010101010101U; >> mask = (mask << 8) - mask; > > (mask << 7) I assume...No. It''s "(mask << 8) - mask". We want to expand the bit at position 0 (in each byte) to the full byte, resulting in 0xFF if the bit is at 1, and 0x00 if the bit is 0. (0 << 8) - 0 = 0x00 (1 << 8) - 1 = 0x100 - 1 = 0xFF Ciao, Andrea -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/20/2013 01:04 PM, Andrea Mazzoleni wrote:> Hi Peter, > >>> static inline uint64_t d2_64(uint64_t v) >>> { >>> uint64_t mask = v & 0x0101010101010101U; >>> mask = (mask << 8) - mask; >> >> (mask << 7) I assume... > No. It''s "(mask << 8) - mask". We want to expand the bit at position 0 > (in each byte) to the full byte, resulting in 0xFF if the bit is at 1, > and 0x00 if the bit is 0. > > (0 << 8) - 0 = 0x00 > (1 << 8) - 1 = 0x100 - 1 = 0xFF >Oh, right... it is the same as (v << 1) - (v >> 7) except everything is shifted over one. -hpa -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Peter,> Now, that doesn''t sound like something that can get neatly meshed into > the Cauchy matrix scheme, I assume.You are correct. Multiplication by 2^-1 cannot be used for the Cauchy method. I used it to implement an alternate triple parity not requiring PSHUFB that I used as reference for performance evaluation of the Cauchy way, assuming that this implementation using powers of 1,2,2^-1 is the fastest possible one. Hopefully the difference is minimal, and the Cauchy method is competitive even at triple parity. Ciao, Andrea -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi,> First, create a 3 by 6 cauchy matrix, using x_i = 2^-i, and y_i = 0 for i=0, and y_i = 2^i for other i. > In this case: x = { 1, 142, 71, 173, 216, 108 } y = { 0, 2, 4). The cauchy matrix is: > > 1 2 4 8 16 32 > 244 83 78 183 118 47 > 167 39 213 59 153 82 > > Divide row 2 by 244 and row 3 by 167. Then extend it with a row of ones on top and it''s still MDS, > and that''s the code for m=4, with RAID-6 as a subset. Very nice!You got it Jim! Thanks, Andrea -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/20/2013 12:30 PM, James Plank wrote:> Peter, I think I understand it differently. Concrete example in GF(256) for k=6, m=4: > > First, create a 3 by 6 cauchy matrix, using x_i = 2^-i, and y_i = 0 for i=0, and y_i = 2^i for other i. In this case: x = { 1, 142, 71, 173, 216, 108 } y = { 0, 2, 4). The cauchy matrix is:Sorry, I took xi and yj to mean a constant x multiplied with i and a constant y multiplied with j, rather than x_i and y_j. -hpa -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Piergiorgio,> In RAID-6 (as per raid6check) there is an easy way > to verify where an HDD has incorrect data. > I suspect, for each 2 parity block it should be > possible to find 1 error (and if this is true, then > quad parity is more attractive than triple one).Yes. The theory say that with quad parity is possible to find 2 errors. Or find 1 error when running in degraded mode with 2 missing disks. But yep, the problem is how to do in a fast way. I don''t know enough the field to answer this. Ciao, Andrea -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Nov 20, 2013 at 11:44:39AM +0100, David Brown wrote: [...]> > In RAID-6 (as per raid6check) there is an easy way > > to verify where an HDD has incorrect data. > > > > I think the way to do that is just to generate the parity blocks from > the data blocks, and compare them to the existing parity blocks.Uhm, the generic RS decoder should try all the possible combination of erasure and so detect the error. This is unfeasible already with 3 parities, so there are faster algorithms, I believe: Peterson–Gorenstein–Zierler algorithm Berlekamp–Massey algorithm Nevertheless, I do not know too much about those, so I cannot state if they apply to the Cauchy matrix as explained here. bye, -- piergiorgio -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Mon, Nov 18, 2013 at 11:08:59PM +0100, Andrea Mazzoleni wrote: [...] I''ve a side question, a bit OT, but maybe you could help with the answer. How about par2? How does this work? They claim "Vendermonde" matrix and they seem to be quite flexible in amount of parities. The could be in GF(2^16), so maybe this makes a difference... bye, -- piergiorgio -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/20/2013 10:16 AM, James Plank wrote:> Hi all -- no real comments, except as I mentioned to Ric, my tutorial > in FAST last February presents Reed-Solomon coding with Cauchy > matrices, and then makes special note of the common pitfall of > assuming that you can append a Vandermonde matrix to an identity > matrix. Please see > http://web.eecs.utk.edu/~plank/plank/papers/2013-02-11-FAST-Tutorial.pdf, > slides 48-52. > > Andrea, does the matrix that you included in an earlier mail (the one > that has Linux RAID-6 in the first two rows) have a general form, or > did you develop it in an ad hoc manner so that it would include Linux > RAID-6 in the first two rows?Hello Jim, It''s always perilous to follow a Ph.D., so I guess I''m feeling suicidal today. ;) I''m not attempting to marginalize Andrea''s work here, but I can''t help but ponder what the real value of triple parity RAID is, or quad, or beyond. Some time ago parity RAID''s primary mission ceased to be surviving single drive failure, or a 2nd failure during rebuild, and became mitigating UREs during a drive rebuild. So we''re now talking about dedicating 3 drives of capacity to avoiding disaster due to platter defects and secondary drive failure. For small arrays this is approaching half the array capacity. So here parity RAID has lost the battle with RAID10''s capacity disadvantage, yet it still suffers the vastly inferior performance in normal read/write IO, not to mention rebuild times that are 3-10x longer. WRT rebuild times, once drives hit 20TB we''re looking at 18 hours just to mirror a drive at full streaming bandwidth, assuming 300MB/s average--and that is probably being kind to the drive makers. With 6 or 8 of these drives, I''d guess a typical md/RAID6 rebuild will take at minimum 72 hours or more, probably over 100, and probably more yet for 3P. And with larger drive count arrays the rebuild times approach a week. Whose users can go a week with degraded performance? This is simply unreasonable, at best. I say it''s completely unacceptable. With these gargantuan drives coming soon, the probability of multiple UREs during rebuild are pretty high. Continuing to use ever more complex parity RAID schemes simply increases rebuild time further. The longer the rebuild, the more likely a subsequent drive failure due to heat buildup, vibration, etc. Thus, in our maniacal efforts to mitigate one failure mode we''re increasing the probability of another. TANSTAFL. Worse yet, RAID10 isn''t going to survive because UREs on a single drive are increasingly likely with these larger drives, and one URE during rebuild destroys the array. I think people are going to have to come to grips with using more and more drives simply to brace the legs holding up their arrays; comes to grips with these insane rebuild times; or bite the bullet they so steadfastly avoided with RAID10. Lots more spindles solves problems, but at a greater cost--again, no free lunch. What I envision is an array type, something similar to RAID 51, i.e. striped parity over mirror pairs. In the case of Linux, this would need to be a new distinct md/RAID level, as both the RAID5 and RAID1 code would need enhancement before being meshed together into this new level[1]. Potential Advantages: 1. Only +1 disk capacity overhead vs RAID 10, regardless of drive count 2. Rebuild time is the same as RAID 10, unless a mirror pair is lost 3. Parity is only used during rebuild if/when a URE occurs, unless ^ 4. Single drive failure doesn''t degrade the parity array, multiple failures in different mirrors doesn''t degrade the parity array 5. Can sustain a minimum of 3 simultaneous drive failures--both drives in one mirror and one drive in another mirror 6. Can lose a maximum of 1/2 of the drives plus 1 drive--one more than RAID 10. Can lose half the drives and still not degrade parity, if no two comprise one mirror 7. Similar or possibly better read throughput vs triple parity RAID 8. Superior write performance with drives down 9. Vastly superior rebuild performance, as rebuilds will rarely, if ever, involve parity Potential Disadvantages: 1. +1 disk overhead vs RAID 10, many more than 2/3P w/large arrays 2. Read-modify-write penalty vs RAID 10 3. Slower write throughput vs triple parity RAID due to spindle deficit 4. Development effort 5. ?? [1] The RAID1/5 code would need to be patched to properly handle a URE encountered by the RAID1 code during rebuild. There are surely other modifications and/or optimizations that would be needed. For large sequential reads, more deterministic read interleaving between mirror pairs would be a good candidate I think. IIUC the RAID1 driver does read interleaving on a per thread basis or some such, which I don''t believe is going to work for this "RAID 51" scenario, at least not for single streaming reads. If this can be done well, we double the read performance of RAID5, and thus we don''t completely "waste" all the extra disks vs big_parity schemes. This proposed "RAID level 51" should have drastically lower rebuild times vs traditional striped parity, should not suffer read/write performance degradation with most disk failure scenarios, and with a read interleaving optimization may have significantly greater streaming read throughput as well. This is far from a perfect solution and I am certainly not promoting it as such. But I think it does have some serious advantages over traditional striped parity schemes, and at minimum is worth discussion as a counterpoint of sorts. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
For myself or any machines I managed for work that do not need high IOPS, I would definitely choose triple- or quad-parity over RAID 51 or similar schemes with arrays of 16 - 32 drives. No need to go into detail here on a subject Adam Leventhal has already covered in detail in an article "Triple-Parity RAID and Beyond" which seems to match the subject of this thread quite nicely: http://queue.acm.org/detail.cfm?id=1670144 -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/20/2013 12:44 PM, Andrea Mazzoleni wrote:> Yes. There are still AMD CPUs sold without SSSE3. Most notably Athlon. > Instead, Intel is providing SSSE3 from the Core 2 Duo.I hate branding discontinuity, due to the resulting confusion... Athlon, Athlon64, Athlon64 X2, Athlon X2 (K10), Athlon II X2, Athlon X2 (Piledriver). Anyone confused? The Trinity and Richland core "Athlon X2" and "Athlon X4" branded processors certainly do support SSSE3, as well as SSE4, AVX, etc. These are the dual/quad core APUs whose graphics cores don''t pass QC and are surgically disabled. AMD decided to brand them as "Athlon" processors. Available since ~2011. For example: http://www.cpu-world.com/CPUs/Bulldozer/AMD-Athlon%20X2%20370K%20-%20AD370KOKA23HL%20-%20AD370KOKHLBOX.html The "Athlon II X2/X3/X4" processors have been out of production for a couple of years now, but a scant few might still be found for sale in the channel. The X2 is based on the clean sheet Regor dual core 45nm design. The X3 and X4 are Phenom II rejects with various numbers of defective cores and defective L3 caches. None support SSSE3. To say "there are still AMD CPUs sold without SSSE3... Most notably Athlon" may be technically true if some Athlon II stragglers exist in the channel. But it isn''t really a fair statement of today''s reality. AMD hasn''t manufactured a CPU without SSSE3 for a couple of years now. And few, if any, Athlon II X2/3/4 chips lacking SSSE3 are for sale. Though there are certainly many such chips still in deployed desktop machines.> A detailed list is available at: http://en.wikipedia.org/wiki/SSSE3Never trust Wikipedia articles to be complete and up to date. However, it does mention Athlon X2 and X4 as planned future product in the Piledriver lineup. Obviously this should be updated to past tense. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/20/2013 8:46 PM, John Williams wrote:> For myself or any machines I managed for work that do not need high > IOPS, I would definitely choose triple- or quad-parity over RAID 51 or > similar schemes with arrays of 16 - 32 drives.You must see a week long rebuild as acceptable...> No need to go into detail hereI disagree.> on a subject Adam Leventhal has already > covered in detail in an article "Triple-Parity RAID and Beyond" which > seems to match the subject of this thread quite nicely: > > http://queue.acm.org/detail.cfm?id=1670144Mr. Leventhal did not address the overwhelming problem we face, which is (multiple) parity array reconstruction time. He assumes the time to simply ''populate'' one drive at its max throughput is the total reconstruction time for the array. While this is typically true for mirror based arrays, it is clearly not for parity arrays. The primary focus of my comments was reducing rebuild time, thus increasing overall reliability. RAID 51 or something similar would achieve this. Thus I think we should discuss alternatives to multiple parity in detail. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Nov 20, 2013 at 10:52 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote:> On 11/20/2013 8:46 PM, John Williams wrote: >> For myself or any machines I managed for work that do not need high >> IOPS, I would definitely choose triple- or quad-parity over RAID 51 or >> similar schemes with arrays of 16 - 32 drives. > > You must see a week long rebuild as acceptable...It would not be a problem if it did take that long, since I would have extra parity units as backup in case of a failure during a rebuild. But of course it would not take that long. Take, for example, a 24 x 3TB triple-parity array (21+3) that has had two drive failures (perhaps the rebuild started with one failure, but there was soon another failure). I would expect the rebuild to take about a day.>> on a subject Adam Leventhal has already >> covered in detail in an article "Triple-Parity RAID and Beyond" which >> seems to match the subject of this thread quite nicely: >> >> http://queue.acm.org/detail.cfm?id=1670144 > > Mr. Leventhal did not address the overwhelming problem we face, which is > (multiple) parity array reconstruction time. He assumes the time to > simply ''populate'' one drive at its max throughput is the total > reconstruction time for the array.Since Adam wrote the code for RAID-Z3 for ZFS, I''m sure he is aware of the time to restore data to failed drives. I do not see any flaw in his analysis related to the time needed to restore data to failed drives. -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 21/11/2013 02:28, Stan Hoeppner wrote:> On 11/20/2013 10:16 AM, James Plank wrote: >> Hi all -- no real comments, except as I mentioned to Ric, my tutorial >> in FAST last February presents Reed-Solomon coding with Cauchy >> matrices, and then makes special note of the common pitfall of >> assuming that you can append a Vandermonde matrix to an identity >> matrix. Please see >> http://web.eecs.utk.edu/~plank/plank/papers/2013-02-11-FAST-Tutorial.pdf, >> slides 48-52. >> >> Andrea, does the matrix that you included in an earlier mail (the one >> that has Linux RAID-6 in the first two rows) have a general form, or >> did you develop it in an ad hoc manner so that it would include Linux >> RAID-6 in the first two rows? > Hello Jim, > > It''s always perilous to follow a Ph.D., so I guess I''m feeling suicidal > today. ;) > > I''m not attempting to marginalize Andrea''s work here, but I can''t help > but ponder what the real value of triple parity RAID is, or quad, or > beyond. Some time ago parity RAID''s primary mission ceased to be > surviving single drive failure, or a 2nd failure during rebuild, and > became mitigating UREs during a drive rebuild. So we''re now talking > about dedicating 3 drives of capacity to avoiding disaster due to > platter defects and secondary drive failure. For small arrays this is > approaching half the array capacity. So here parity RAID has lost the > battle with RAID10''s capacity disadvantage, yet it still suffers the > vastly inferior performance in normal read/write IO, not to mention > rebuild times that are 3-10x longer. > > WRT rebuild times, once drives hit 20TB we''re looking at 18 hours just > to mirror a drive at full streaming bandwidth, assuming 300MB/s > average--and that is probably being kind to the drive makers. With 6 or > 8 of these drives, I''d guess a typical md/RAID6 rebuild will take at > minimum 72 hours or more, probably over 100, and probably more yet for > 3P. And with larger drive count arrays the rebuild times approach a > week. Whose users can go a week with degraded performance? This is > simply unreasonable, at best. I say it''s completely unacceptable. > > With these gargantuan drives coming soon, the probability of multiple > UREs during rebuild are pretty high.No because if you are correct about the very high CPU overhead during rebuild (which I don''t see so dramatic as Andrea claims 500MB/sec for triple-parity, probably parallelizable on multiple cores), the speed of rebuild decreases proportionally and hence the stress and heating on the drives proportionally reduces, approximating that of normal operation. And how often have you seen a drive failure in a week during normal operation? But in reality, consider that a non-naive implementation of multiple-parity would probably use just the single parity during reconstruction if just one disk fails, using the multiple parities only to read the stripes which are unreadable at single parity. So the speed and time of reconstruction and performance penalty would be that of raid5 except in exceptional situations of multiple failures.> ... > What I envision is an array type, something similar to RAID 51, i.e. > striped parity over mirror pairs. ....I don''t like your approach of raid 51: it has the write overhead of raid5, with the waste of space of raid1. So it cannot be used as neither a performance array nor a capacity array. In the scope of this discussion (we are talking about very large arrays), the waste of space of your solution, higher than 50%, will make your solution costing double the price. A competitor for the multiple-parity scheme might be raid65 or 66, but this is a so much dirtier approach than multiple parity if you think at the kind of rmw and overhead that will occur during normal operation. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 20/11/13 19:09, John Williams wrote:> On Wed, Nov 20, 2013 at 2:31 AM, David Brown <david.brown@hesbynett.no> wrote: >> That''s certainly a reasonable way to look at it. We should not limit >> the possibilities for high-end systems because of the limitations of >> low-end systems that are unlikely to use 3+ parity anyway. I''ve also >> looked up a list of the processors that support SSE3 and PSHUFB - a lot >> of modern "low-end" x86 cpus support it. And of course it is possible >> to implement general G(2^8) multiplication without PSHUFB, using a >> lookup table - it is important that this can all work with any CPU, even >> if it is slow. > > Unfortunately, it is SSSE3 that is required for PSHUFB. The SSE3 set > with only two-esses does not suffice. I made that same mistake when I > first heard about Andrea''s 6-parity work. SSSE3 vs. SSE3, confusing > notation! > > SSSE3 is significantly less widely supported than SSE3. Particularly > on AMD, only the very latest CPUs seem to support SSSE3. Intel support > for SSSE3 goes back much further than AMD support. > > Maybe it is not such a big problem, since it may be possible to > support two "roads". Both roads would include the current md RAID-5 > and RAID-6. But one road, which those lacking CPUs supporting SSSE3 > might choose, would continue on to the non-SSSE3 triple-parity 2^-1 > technique, and then dead-end. The other road would continue with the > Cauchy matrix technique through 3-parity all the way to 6-parity. > > It might even be feasible to allow someone stuck at the end of the > non-SSSE3 road to convert to the Cauchy road. You would have to go > through all the 2^-1 triple-parity and convert it to Cauchy > triple-parity. But then you would be safely on the Cauchy road. >I would not like to see two alternative triple-parity solutions - I think that would lead to confusion, and a non-Cauchy triple parity would not be extendible without a rebuild (I''ve talked before about the idea of temporarily adding an extra parity drive with an asymmetric layout. I really like the idea, so I keep pushing for it!). I think it is better to accept that 3+ parity will be slow on processors that don''t support PSHUFB. We should try to find the best alternative SIMD for other realistic processors (such as on AMD chips without PSHUFB, ARM''s with NEON, PPC with Altivec, etc.) - but a simple table lookup will always work as a fallback. Other than that I think it is fair to say that if you want /fast/ 3+ parity, you need a reasonably modern non-budget-class cpu. -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 20/11/13 19:34, Andrea Mazzoleni wrote:> Hi David, > >>> The choice of ZFS to use powers of 4 was likely not optimal, >>> because to multiply by 4, it has to do two multiplications by 2. >> I can agree with that. I didn''t copy ZFS''s choice here > David, it was not my intention to suggest that you copied from ZFS. > Sorry to have expressed myself badly. I just mentioned ZFS because it''s > an implementation that I know uses powers of 4 to generate triple > parity, and I saw in the code that it''s implemented with two multiplication > by 2. >Andrea, I didn''t take your comment as an accusation of any kind - there is no need for any kind of apology! It was was merely a statement of fact - I picked powers of 4 as an obvious extension of the powers of 2 in raid6, and found it worked well. And of course, in the open source world, copying of code and ideas is a good thing - there is no point in re-inventing the wheel unless we can invent a better one. Really, I /should/ have read the ZFS implementation and copied it! mvh., David -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 21/11/13 02:28, Stan Hoeppner wrote:> On 11/20/2013 10:16 AM, James Plank wrote: >> Hi all -- no real comments, except as I mentioned to Ric, my tutorial >> in FAST last February presents Reed-Solomon coding with Cauchy >> matrices, and then makes special note of the common pitfall of >> assuming that you can append a Vandermonde matrix to an identity >> matrix. Please see >> http://web.eecs.utk.edu/~plank/plank/papers/2013-02-11-FAST-Tutorial.pdf, >> slides 48-52. >> >> Andrea, does the matrix that you included in an earlier mail (the one >> that has Linux RAID-6 in the first two rows) have a general form, or >> did you develop it in an ad hoc manner so that it would include Linux >> RAID-6 in the first two rows? > > Hello Jim, > > It''s always perilous to follow a Ph.D., so I guess I''m feeling suicidal > today. ;) > > I''m not attempting to marginalize Andrea''s work here, but I can''t help > but ponder what the real value of triple parity RAID is, or quad, or > beyond. Some time ago parity RAID''s primary mission ceased to be > surviving single drive failure, or a 2nd failure during rebuild, and > became mitigating UREs during a drive rebuild. So we''re now talking > about dedicating 3 drives of capacity to avoiding disaster due to > platter defects and secondary drive failure. For small arrays this is > approaching half the array capacity. So here parity RAID has lost the > battle with RAID10''s capacity disadvantage, yet it still suffers the > vastly inferior performance in normal read/write IO, not to mention > rebuild times that are 3-10x longer. > > WRT rebuild times, once drives hit 20TB we''re looking at 18 hours just > to mirror a drive at full streaming bandwidth, assuming 300MB/s > average--and that is probably being kind to the drive makers. With 6 or > 8 of these drives, I''d guess a typical md/RAID6 rebuild will take at > minimum 72 hours or more, probably over 100, and probably more yet for > 3P. And with larger drive count arrays the rebuild times approach a > week. Whose users can go a week with degraded performance? This is > simply unreasonable, at best. I say it''s completely unacceptable. > > With these gargantuan drives coming soon, the probability of multiple > UREs during rebuild are pretty high. Continuing to use ever more > complex parity RAID schemes simply increases rebuild time further. The > longer the rebuild, the more likely a subsequent drive failure due to > heat buildup, vibration, etc. Thus, in our maniacal efforts to mitigate > one failure mode we''re increasing the probability of another. TANSTAFL. > Worse yet, RAID10 isn''t going to survive because UREs on a single drive > are increasingly likely with these larger drives, and one URE during > rebuild destroys the array. >I don''t think the chances of hitting an URE during rebuild is dependent on the rebuild time - merely on the amount of data read during rebuild. URE rates are "per byte read" rather than "per unit time", are they not? I think you are overestimating the rebuild times a bit, but there is no arguing that rebuild on parity raids is a lot more work (for the cpu, the IO system, and the disks) than for mirror raids.> I think people are going to have to come to grips with using more and > more drives simply to brace the legs holding up their arrays; comes to > grips with these insane rebuild times; or bite the bullet they so > steadfastly avoided with RAID10. Lots more spindles solves problems, > but at a greater cost--again, no free lunch. > > What I envision is an array type, something similar to RAID 51, i.e. > striped parity over mirror pairs. In the case of Linux, this would need > to be a new distinct md/RAID level, as both the RAID5 and RAID1 code > would need enhancement before being meshed together into this new level[1].Shouldn''t we be talking about RAID 15 here, rather than RAID 51 ? I interpret "RAID 15" to be like "RAID 10" - a raid5 set of raid1 mirrors, while "RAID 51" would be a raid1 mirror of raid5 sets. I am certain that you mean a raid5 set of raid1 pairs - I just think you''ve got the name wrong.> > Potential Advantages: > > 1. Only +1 disk capacity overhead vs RAID 10, regardless of drive count+2 disks (the raid5 parity "disk" is a raid1 pair)> 2. Rebuild time is the same as RAID 10, unless a mirror pair is lost > 3. Parity is only used during rebuild if/when a URE occurs, unless ^ > 4. Single drive failure doesn''t degrade the parity array, multiple > failures in different mirrors doesn''t degrade the parity array > 5. Can sustain a minimum of 3 simultaneous drive failures--both drives > in one mirror and one drive in another mirror > 6. Can lose a maximum of 1/2 of the drives plus 1 drive--one more than > RAID 10. Can lose half the drives and still not degrade parity, > if no two comprise one mirror > 7. Similar or possibly better read throughput vs triple parity RAID > 8. Superior write performance with drives down > 9. Vastly superior rebuild performance, as rebuilds will rarely, if > ever, involve parity > > Potential Disadvantages: > > 1. +1 disk overhead vs RAID 10, many more than 2/3P w/large arrays > 2. Read-modify-write penalty vs RAID 10 > 3. Slower write throughput vs triple parity RAID due to spindle deficit > 4. Development effort > 5. ?? > > > [1] The RAID1/5 code would need to be patched to properly handle a URE > encountered by the RAID1 code during rebuild. There are surely other > modifications and/or optimizations that would be needed. For large > sequential reads, more deterministic read interleaving between mirror > pairs would be a good candidate I think. IIUC the RAID1 driver does > read interleaving on a per thread basis or some such, which I don''t > believe is going to work for this "RAID 51" scenario, at least not for > single streaming reads. If this can be done well, we double the read > performance of RAID5, and thus we don''t completely "waste" all the extra > disks vs big_parity schemes. > > This proposed "RAID level 51" should have drastically lower rebuild > times vs traditional striped parity, should not suffer read/write > performance degradation with most disk failure scenarios, and with a > read interleaving optimization may have significantly greater streaming > read throughput as well. > > This is far from a perfect solution and I am certainly not promoting it > as such. But I think it does have some serious advantages over > traditional striped parity schemes, and at minimum is worth discussion > as a counterpoint of sorts. >I don''t see that there needs to be any changes to the existing md code to make raid15 work - it is merely a raid 5 made from a set of raid1 pairs. I can see that improved threading and interleaving could be a benefit here - but that''s the case in general for md raid, and it is something that the developers are already working on (I haven''t followed the details, but the topic comes up regularly on the list here). So as far as I can see, you''ve got raid15 support already - if that''s what suits your needs, use it. Future improvements to the md code are only needed to make it faster. Of course, there is scope for making specific raid15 support in md along the lines of the raid10 code - raid15,f2 would have the same speed advantages over "normal" raid1+5 as raid10,f2 has over raid1+0. Whether it is worth the effort implementing it is a different matter. I can see plenty of reasons why raid15 might be a good idea, and even raid16 for 5 disk redundancy, compared to multi-parity sets. However, it costs a lot in disk space. For example, with 20 disks at 1 TB each, you can have: raid5 = 19TB, 1 disk redundancy raid6 = 18TB, 2 disk redundancy raid6.3 = 17TB, 3 disk redundancy raid6.4 = 16TB, 4 disk redundancy raid6.5 = 15TB, 5 disk redundancy raid10 = 10TB, 1 disk redundancy raid15 = 8TB, 3 disk redundancy raid16 = 6TB, 5 disk redundancy That''s a very significant difference. Implementing 3+ parity does not stop people using raid15, or similar schemes - it just adds more choice to let people optimise according to their needs. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 21/11/13 20:07, David Brown wrote:> I can see plenty of reasons why raid15 might be a good idea, and even > raid16 for 5 disk redundancy, compared to multi-parity sets. However, > it costs a lot in disk space. For example, with 20 disks at 1 TB each, > you can have: > > raid5 = 19TB, 1 disk redundancy > raid6 = 18TB, 2 disk redundancy > raid6.3 = 17TB, 3 disk redundancy > raid6.4 = 16TB, 4 disk redundancy > raid6.5 = 15TB, 5 disk redundancy > > raid10 = 10TB, 1 disk redundancy > raid15 = 8TB, 3 disk redundancy > raid16 = 6TB, 5 disk redundancy > > > That''s a very significant difference. > > Implementing 3+ parity does not stop people using raid15, or similar > schemes - it just adds more choice to let people optimise according to > their needs.BTW, as far as strange RAID type options to try and get around problems with failed disks, before I learned about timeout mismatches, I was pretty worried when my 5 disk RAID5 kept falling apart and losing a random member, then adding the failed disk back would work perfectly. To help me feel better about this, I used 5 x 500GB drives in RAID5 and then used the RAID5 + 1 x 2TB drive in RAID1, meaning I could afford to lose any two disks without losing data. Of course, now I know RAID6 might have been a better choice, or even simply 2 x 2TB drives in RAID1 :) In any case, I''m not sure I understand the concern with RAID 7.X (as it is being called, where X > 2). Certainly you will need to make 1 computation for each stripe being written, for each value of X, so RAID 7.5 with 5 disk redundancy means 5 calculations for each stripe being written. However, given that drives are getting bigger every year, did we forget that we are also getting faster CPU and also more cores in a single "CPU package"? On a pure storage server, the CPU would normally have nothing to do, except a little interrupt handling, it is just shuffling bytes around. Of course, if you need RAID7.5 then you probably have a dedicated storage server, so I don''t see the problem with using the CPU to do all the calculations. Of course, if you are asking about carbon emissions, and cooling costs in the data center, this could (on a global scale) have a significant impact, so maybe it is a bad idea after all :) Regards, Adam -- Adam Goryachev Website Managers www.websitemanagers.com.au -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 20/11/13 22:59, Piergiorgio Sartor wrote:> On Wed, Nov 20, 2013 at 11:44:39AM +0100, David Brown wrote: > [...] >>> In RAID-6 (as per raid6check) there is an easy way >>> to verify where an HDD has incorrect data. >>> >> >> I think the way to do that is just to generate the parity blocks from >> the data blocks, and compare them to the existing parity blocks. > > Uhm, the generic RS decoder should try all > the possible combination of erasure and so > detect the error. > This is unfeasible already with 3 parities, > so there are faster algorithms, I believe: > > Peterson–Gorenstein–Zierler algorithm > Berlekamp–Massey algorithm > > Nevertheless, I do not know too much about > those, so I cannot state if they apply to > the Cauchy matrix as explained here. > > bye, >Ah, you are trying to find which disk has incorrect data so that you can change just that one disk? There are dangers with that... <http://neil.brown.name/blog/20100211050355> If you disagree with this blog post (and I urge you to read it in full first), then this is how I would do a "smart" stripe recovery: First calculate the parities from the data blocks, and compare these with the existing parity blocks. If they all match, the stripe is consistent. Normal (detectable) disk errors and unrecoverable read errors get flagged by the disk and the IO system, and you /know/ there is a problem with that block. Whether it is a data block or a parity block, you re-generate the correct data and store it - that''s what your raid is for. If you have no detected read errors, and there is one parity inconsistency, then /probably/ that block has had an undetected read error, or it simply has not been written completely before a crash. Either way, just re-write the correct parity. If there are two or more parity inconsistencies, but not all parities are in error, then you either have multiple disk or block failures, or you have a partly-written stripe. Any attempts at "smart" correction will almost certainly be worse than just re-writing the new parities and hoping that the filesystem''s journal works. If all the parities are inconsistent, then the "smart" thing is to look for a single incorrect disk block. Just step through the blocks one by one - assume that block is wrong and replace it (in temporary memory, not on disk!) with a recovered version from the other data blocks and the parities (only the first parity is needed). Re-calculate the other parities and compare. If the other parities now match, then you have found a single inconsistent data block. It /may/ be a good idea to re-write this - or maybe not (see the blog post linked above). If you don''t find any single data blocks that can be "corrected" in this way, then re-writing the parity blocks to match the disk data is probably the least harmful fix. Remember, this is not a general error detection and correction scheme - it is a system targeted for a particular type of use, with particular patterns of failure and failure causes, and particular mechanisms on top (journalled file systems) to consider. -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 21/11/13 10:54, Adam Goryachev wrote:> On 21/11/13 20:07, David Brown wrote: >> I can see plenty of reasons why raid15 might be a good idea, and even >> raid16 for 5 disk redundancy, compared to multi-parity sets. However, >> it costs a lot in disk space. For example, with 20 disks at 1 TB each, >> you can have: >> >> raid5 = 19TB, 1 disk redundancy >> raid6 = 18TB, 2 disk redundancy >> raid6.3 = 17TB, 3 disk redundancy >> raid6.4 = 16TB, 4 disk redundancy >> raid6.5 = 15TB, 5 disk redundancy >> >> raid10 = 10TB, 1 disk redundancy >> raid15 = 8TB, 3 disk redundancy >> raid16 = 6TB, 5 disk redundancy >> >> >> That''s a very significant difference. >> >> Implementing 3+ parity does not stop people using raid15, or similar >> schemes - it just adds more choice to let people optimise according to >> their needs. > BTW, as far as strange RAID type options to try and get around problems > with failed disks, before I learned about timeout mismatches, I was > pretty worried when my 5 disk RAID5 kept falling apart and losing a > random member, then adding the failed disk back would work perfectly. To > help me feel better about this, I used 5 x 500GB drives in RAID5 and > then used the RAID5 + 1 x 2TB drive in RAID1, meaning I could afford to > lose any two disks without losing data. Of course, now I know RAID6 > might have been a better choice, or even simply 2 x 2TB drives in RAID1 :) > > In any case, I''m not sure I understand the concern with RAID 7.X (as it > is being called, where X > 2). Certainly you will need to make 1 > computation for each stripe being written, for each value of X, so RAID > 7.5 with 5 disk redundancy means 5 calculations for each stripe being > written. However, given that drives are getting bigger every year, did > we forget that we are also getting faster CPU and also more cores in a > single "CPU package"? >This is all true. And md code is getting better at using more cores under more circumstances, making the parity calculations more efficient. The speed concern (which was Stan''s, rather than mine) is more about recovery and rebuild. If you have a layered raid with raid1 pairs at the bottom level, then recovery and rebuild (from a single failure) is just a straight copy from one disk to another - you don''t get faster than that. If you have a 20 + 3 parity raid, then rebuilding requires reading a stripe from 20 disks and writing to 1 disk - that''s far more effort and is likely to take more time unless your IO system can handle full bandwidth of all the disks simultaneously. Similarly, performance of the array while rebuilding or degraded is much worse for parity raids than for raids on top of raid1 pairs. How that matters to you, and how it balances with the space costs, is up to you and your application.> On a pure storage server, the CPU would normally have nothing to do, > except a little interrupt handling, it is just shuffling bytes around. > Of course, if you need RAID7.5 then you probably have a dedicated > storage server, so I don''t see the problem with using the CPU to do all > the calculations. > > Of course, if you are asking about carbon emissions, and cooling costs > in the data center, this could (on a global scale) have a significant > impact, so maybe it is a bad idea after all :) > > Regards, > Adam >-- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 2013-11-21 11:13, David Brown wrote:> On 20/11/13 22:59, Piergiorgio Sartor wrote: >> On Wed, Nov 20, 2013 at 11:44:39AM +0100, David Brown wrote: >> [...] >>>> In RAID-6 (as per raid6check) there is an easy way >>>> to verify where an HDD has incorrect data. >>>> >>> >>> I think the way to do that is just to generate the parity blocks from >>> the data blocks, and compare them to the existing parity blocks. >> >> Uhm, the generic RS decoder should try all >> the possible combination of erasure and so >> detect the error. >> This is unfeasible already with 3 parities, >> so there are faster algorithms, I believe: >> >> Peterson–Gorenstein–Zierler algorithm >> Berlekamp–Massey algorithm >> >> Nevertheless, I do not know too much about >> those, so I cannot state if they apply to >> the Cauchy matrix as explained here. >> >> bye, >> > > Ah, you are trying to find which disk has incorrect data so that you can > change just that one disk? There are dangers with that... > > <http://neil.brown.name/blog/20100211050355> > > If you disagree with this blog post (and I urge you to read it in full > first), then this is how I would do a "smart" stripe recovery: > > > First calculate the parities from the data blocks, and compare these > with the existing parity blocks. > > If they all match, the stripe is consistent. > > Normal (detectable) disk errors and unrecoverable read errors get > flagged by the disk and the IO system, and you /know/ there is a problem > with that block. Whether it is a data block or a parity block, you > re-generate the correct data and store it - that''s what your raid is for. > > If you have no detected read errors, and there is one parity > inconsistency, then /probably/ that block has had an undetected read > error, or it simply has not been written completely before a crash. > Either way, just re-write the correct parity. > > If there are two or more parity inconsistencies, but not all parities > are in error, then you either have multiple disk or block failures, or > you have a partly-written stripe. Any attempts at "smart" correction > will almost certainly be worse than just re-writing the new parities and > hoping that the filesystem''s journal works. > > If all the parities are inconsistent, then the "smart" thing is to look > for a single incorrect disk block. Just step through the blocks one by > one - assume that block is wrong and replace it (in temporary memory, > not on disk!) with a recovered version from the other data blocks and > the parities (only the first parity is needed). Re-calculate the other > parities and compare. If the other parities now match, then you have > found a single inconsistent data block. It /may/ be a good idea to > re-write this - or maybe not (see the blog post linked above). > > If you don''t find any single data blocks that can be "corrected" in this > way, then re-writing the parity blocks to match the disk data is > probably the least harmful fix.It has to be pointed out that all filesystems or are trying to integrate or have integrated some sort of checksumming to avoid guessing which between the data and/or the parity is wrong: - btrfs is fully checksummed - zfs is fully checksummed - ext4 and xfs are trying to checksumming the metadata [1][2] - refs (Windows) protects with checksum the metadata and (optionally) the data [3] We are talking about [1] https://ext4.wiki.kernel.org/index.php/Ext4_Metadata_Checksums [2] http://xfs.org/images/d/d1/Xfs-scalability-lca2012.pdf [3] http://en.wikipedia.org/wiki/ReFS> > > Remember, this is not a general error detection and correction scheme - > it is a system targeted for a particular type of use, with particular > patterns of failure and failure causes, and particular mechanisms on top > (journalled file systems) to consider. > > > -- > To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html >-- gpg @keyserver.linux.it: Goffredo Baroncelli (kreijackATinwind.it> Key fingerprint BBF5 1610 0B64 DAC6 5F7D 17B2 0EDA 9B37 8B82 E0B5 -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Nov 20, 2013 at 07:28:37PM -0600, Stan Hoeppner wrote: [...]> It''s always perilous to follow a Ph.D., so I guess I''m feeling suicidal > today. ;) > > I''m not attempting to marginalize Andrea''s work here, but I can''t help > but ponder what the real value of triple parity RAID is, or quad, or > beyond. Some time ago parity RAID''s primary mission ceased to beHi Stan, my opinio is that you have to think in terms of storage devices which are not always available. Those are not simply directly connected HDDs, it could be more exotic. The example I consider is a p2p network storage, where the nodes are very little reliable. I guess that could be more. bye, -- piergiorgio -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Nov 21, 2013 at 11:13:29AM +0100, David Brown wrote: [...]> Ah, you are trying to find which disk has incorrect data so that you can > change just that one disk? There are dangers with that...Hi David,> <http://neil.brown.name/blog/20100211050355>I think we already did the exercise, here :-)> If you disagree with this blog post (and I urge you to read it in fullWe discussed the topic (with Neil) and, if I recall correctly, he is agaist having an _automatic_ error detectio and correction _in_ kernel. I fully agree with that: user space is better and it should not be automatic, but it should do things under user control. The current "check" operetion is pretty poor. It just reports how many mismatches, it does not even report where in the array. The first step, independent from how many parities one has, would be to tell the user where the mismatches occurred, so it would be possible to check the FS at that position. Having a multi parity RAID allows to check even which disk. This would provide the user with a more comprehensive (I forgot the spelling) information. Of course, since we are there, we can also give the option to fix it. This would be much likely a "fsck".> first), then this is how I would do a "smart" stripe recovery: > > First calculate the parities from the data blocks, and compare these > with the existing parity blocks. > > If they all match, the stripe is consistent. > > Normal (detectable) disk errors and unrecoverable read errors get > flagged by the disk and the IO system, and you /know/ there is a problem > with that block. Whether it is a data block or a parity block, you > re-generate the correct data and store it - that''s what your raid is for.That''s not always the case, otherwise having the mismatch count would be useless. The issue is that errors appear, whatever the reason, without being reported by the underlying hardware.> If you have no detected read errors, and there is one parity > inconsistency, then /probably/ that block has had an undetected read > error, or it simply has not been written completely before a crash. > Either way, just re-write the correct parity.Why re-write the parity if I can get the correct data there? If can be sure that one data block is incorrect and I can re-create properly, that''s the thing to do.> Remember, this is not a general error detection and correction scheme -It is not, but it could be. For free. bye, -- piergiorgio -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 21/11/13 21:05, Piergiorgio Sartor wrote:> On Thu, Nov 21, 2013 at 11:13:29AM +0100, David Brown wrote: > [...] >> Ah, you are trying to find which disk has incorrect data so that you can >> change just that one disk? There are dangers with that... > > Hi David, > >> <http://neil.brown.name/blog/20100211050355> > > I think we already did the exercise, here :-) > >> If you disagree with this blog post (and I urge you to read it in full > > We discussed the topic (with Neil) and, if I > recall correctly, he is agaist having an > _automatic_ error detectio and correction _in_ > kernel. > I fully agree with that: user space is better > and it should not be automatic, but it should > do things under user control. >OK.> The current "check" operetion is pretty poor. > It just reports how many mismatches, it does > not even report where in the array. > The first step, independent from how many > parities one has, would be to tell the user > where the mismatches occurred, so it would > be possible to check the FS at that position.Certainly it would be good to give the user more information. If you can tell the user where the errors are, and what the likely failed block is, then that would be very useful. If you can tell where it is in the filesystem (such as which file, if any, owns the blocks in question) then that would be even better.> Having a multi parity RAID allows to check > even which disk. > This would provide the user with a more > comprehensive (I forgot the spelling) > information. > > Of course, since we are there, we can > also give the option to fix it. > This would be much likely a "fsck".If this can all be done to give the user an informed choice, then it sounds good. One issue here is whether the check should be done with the filesystem mounted and in use, or only off-line. If it is off-line then it will mean a long down-time while the array is checked - but if it is online, then there is the risk of confusing the filesystem and caches by changing the data.> >> first), then this is how I would do a "smart" stripe recovery: >> >> First calculate the parities from the data blocks, and compare these >> with the existing parity blocks. >> >> If they all match, the stripe is consistent. >> >> Normal (detectable) disk errors and unrecoverable read errors get >> flagged by the disk and the IO system, and you /know/ there is a problem >> with that block. Whether it is a data block or a parity block, you >> re-generate the correct data and store it - that''s what your raid is for. > > That''s not always the case, otherwise > having the mismatch count would be useless. > The issue is that errors appear, whatever > the reason, without being reported by the > underlying hardware. >(I know you know how this works, so I am not trying to be patronising with this explanation - I just think we have slightly misunderstood what the other is saying, so spelling it out will hopefully make it clearer.) Most disk errors /are/ detectable, and are reported by the underlying hardware - small surface errors are corrected by the disk''s own error checking and correcting mechanisms, and larger errors are usually detected. It is (or should be!) very rare that a read error goes undetected without there being a major problem with the disk controller. And if the error is detected, then the normal raid processing kicks in as there is no doubt about which block has problems.>> If you have no detected read errors, and there is one parity >> inconsistency, then /probably/ that block has had an undetected read >> error, or it simply has not been written completely before a crash. >> Either way, just re-write the correct parity. > > Why re-write the parity if I can get > the correct data there? > If can be sure that one data block is > incorrect and I can re-create properly, > that''s the thing to do.If you can be /sure/ about which data block is incorrect, then I agree - but you can''t be /entirely/ sure. But I agree that you can make a good enough guess to recommend a fix to the user - as long as it is not automatic.> >> Remember, this is not a general error detection and correction scheme - > > It is not, but it could be. For free. >For most ECC schemes, you know that all your blocks are set synchronously - so any block that does not fit in, is an error. With raid, it could also be that a stripe is only partly written - you can have two different valid sets of data mixed to give an inconsistent stripe, without any good way of telling what consistent data is the best choice. Perhaps a checking tool can take advantage of a write-intent bitmap (if there is one) so that it knows if an inconsistent stripe is partly updated or the result of a disk error. mvh., David -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi David, On Thu, Nov 21, 2013 at 09:31:46PM +0100, David Brown wrote: [...]> If this can all be done to give the user an informed choice, then it > sounds good.that would be my target. To _offer_ more options to the (advanced) user. It _must_ always be under user control.> One issue here is whether the check should be done with the filesystem > mounted and in use, or only off-line. If it is off-line then it will > mean a long down-time while the array is checked - but if it is online, > then there is the risk of confusing the filesystem and caches by > changing the data.Currently, "raid6check" can work with FS mounted. I got the suggestion from Neil (of course). It is possible to lock one stripe and check it. This should be, at any given time, consistent (that is, the parity should always match the data). If an error is found, it is reported. Again, the user can decide to fix it or not, considering all the FS consequences and so on.> Most disk errors /are/ detectable, and are reported by the underlying > hardware - small surface errors are corrected by the disk''s own error > checking and correcting mechanisms, and larger errors are usually > detected. It is (or should be!) very rare that a read error goes > undetected without there being a major problem with the disk controller. > And if the error is detected, then the normal raid processing kicks in > as there is no doubt about which block has problems.That''s clear. That case is an "erasure" (I think) and it is perfectly in line with the usual operation. I''m not trying to replace this mechanism.> If you can be /sure/ about which data block is incorrect, then I agree - > but you can''t be /entirely/ sure. But I agree that you can make a good > enough guess to recommend a fix to the user - as long as it is not > automatic.One typical case is when many errors are found, belonging to the same disk. This case clearly shows the disk is to be replaced or the interface checked... But, again, the user is the master, not the machine... :-)> For most ECC schemes, you know that all your blocks are set > synchronously - so any block that does not fit in, is an error. With > raid, it could also be that a stripe is only partly written - you canCould it be? I would consider this an error. The stripe must always be consistent, there should be a transactional mechanism to make sure that, if read back, the data is always matching the parity. When I write "read back" I mean from whatever the data is: physical disk or cache. Otherwise, the check must run exclusively on the array (no mounted FS, no other things running on it).> have two different valid sets of data mixed to give an inconsistent > stripe, without any good way of telling what consistent data is the best > choice. > > Perhaps a checking tool can take advantage of a write-intent bitmap (if > there is one) so that it knows if an inconsistent stripe is partly > updated or the result of a disk error.Of course, this is an option, which should be taken into consideration. Any improvement idea is welcome!!! bye, -- piergiorgio -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/21/2013 1:05 AM, John Williams wrote:> On Wed, Nov 20, 2013 at 10:52 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote: >> On 11/20/2013 8:46 PM, John Williams wrote: >>> For myself or any machines I managed for work that do not need high >>> IOPS, I would definitely choose triple- or quad-parity over RAID 51 or >>> similar schemes with arrays of 16 - 32 drives. >> >> You must see a week long rebuild as acceptable... > > It would not be a problem if it did take that long, since I would have > extra parity units as backup in case of a failure during a rebuild. > > But of course it would not take that long. Take, for example, a 24 x > 3TB triple-parity array (21+3) that has had two drive failures > (perhaps the rebuild started with one failure, but there was soon > another failure). I would expect the rebuild to take about a day.You''re looking at today. We''re discussing tomorrow''s needs. Today''s 6TB 3.5" drives have sustained average throughput of ~175MB/s. Tomorrow''s 20TB drives will be lucky to do 300MB/s. As I said previously, at that rate a straight disk-disk copy of a 20TB drive takes 18.6 hours. This is what you get with RAID1/10/51. In the real world, rebuilding a failed drive in a 3P array of say 8 of these disks will likely take at least 3 times as long, 2 days 6 hours minimum, probably more. This may be perfectly acceptable to some, but probably not to all.>>> on a subject Adam Leventhal has already >>> covered in detail in an article "Triple-Parity RAID and Beyond" which >>> seems to match the subject of this thread quite nicely: >>> >>> http://queue.acm.org/detail.cfm?id=1670144 >> >> Mr. Leventhal did not address the overwhelming problem we face, which is >> (multiple) parity array reconstruction time. He assumes the time to >> simply ''populate'' one drive at its max throughput is the total >> reconstruction time for the array. > > Since Adam wrote the code for RAID-Z3 for ZFS, I''m sure he is aware of > the time to restore data to failed drives. I do not see any flaw in > his analysis related to the time needed to restore data to failed > drives.He wrote that article in late 2009. It seems pretty clear he wasn''t looking 10 years forward to 20TB drives, where the minimum mirror rebuild time will be ~18 hours, and parity rebuild will be much greater. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Nov 21, 2013 at 2:57 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote:> He wrote that article in late 2009. It seems pretty clear he wasn''t > looking 10 years forward to 20TB drives, where the minimum mirror > rebuild time will be ~18 hours, and parity rebuild will be much greater.Actually, it is completely obvious that he WAS looking ten years ahead, seeing as several of his graphs have time scales going to 2009+10 = 2019. And he specifically mentions longer rebuild times as one of the reasons why higher parity RAIDs are needed. -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/21/2013 2:08 AM, joystick wrote:> On 21/11/2013 02:28, Stan Hoeppner wrote:...>> WRT rebuild times, once drives hit 20TB we''re looking at 18 hours just >> to mirror a drive at full streaming bandwidth, assuming 300MB/s >> average--and that is probably being kind to the drive makers. With 6 or >> 8 of these drives, I''d guess a typical md/RAID6 rebuild will take at >> minimum 72 hours or more, probably over 100, and probably more yet for >> 3P. And with larger drive count arrays the rebuild times approach a >> week. Whose users can go a week with degraded performance? This is >> simply unreasonable, at best. I say it''s completely unacceptable. >> >> With these gargantuan drives coming soon, the probability of multiple >> UREs during rebuild are pretty high. > > No because if you are correct about the very high CPU overhead duringI made no such claim.> rebuild (which I don''t see so dramatic as Andrea claims 500MB/sec for > triple-parity, probably parallelizable on multiple cores), the speed of > rebuild decreases proportionallyThe rebuild time of a parity array normally has little to do with CPU overhead. The bulk of the elapsed time is due to: 1. The serial nature of the rebuild algorithm 2. The random IO pattern of the reads 3. The rotational latency of the drives #3 is typically the largest portion of the elapsed time.> and hence the stress and heating on the > drives proportionally reduces, approximating that of normal operation. > And how often have you seen a drive failure in a week during normal > operation?This depends greatly on one''s normal operation. In general, for most users of parity arrays, any full array operation such as a rebuild or reshape is far more taxing on the drives, in both power draw and heat dissipation, than ''normal'' operation.> But in reality, consider that a non-naive implementation of > multiple-parity would probably use just the single parity during > reconstruction if just one disk fails, using the multiple parities only > to read the stripes which are unreadable at single parity. So the speed > and time of reconstruction and performance penalty would be that of > raid5 except in exceptional situations of multiple failures.That may very well be, but it doesn''t change #2,3 above.>> What I envision is an array type, something similar to RAID 51, i.e. >> striped parity over mirror pairs. .... > > I don''t like your approach of raid 51: it has the write overhead of > raid5, with the waste of space of raid1. > So it cannot be used as neither a performance array nor a capacity array.I don''t like it either. It''s a compromise. But as RAID1/10 will soon be unusable due to URE probability during rebuild, I think it''s a relatively good compromise for some users, some workloads.> In the scope of this discussion (we are talking about very large > arrays),Capacity yes, drive count, no. Drive capacities are increasing at a much faster rate than our need for storage space. As we move forward the trend will be building larger capacity arrays with fewer disks.> the waste of space of your solution, higher than 50%, will make > your solution costing double the price.This is the classic mirror vs parity argument. Using 1 more disk to add parity to striped mirrors doesn''t change it. "Waste" is in the eye of the beholder. Anyone currently using RAID10 will have no problem dedicating one more disk for uptime, protection.> A competitor for the multiple-parity scheme might be raid65 or 66, but > this is a so much dirtier approach than multiple parity if you think at > the kind of rmw and overhead that will occur during normal operation.Neither of those has any advantage over multi-parity. I suggested this approach because it retains all of the advantages of RAID10 but one. We sacrifice fast random write performance for protection against UREs, the same reason behind 3P. That''s what the single parity is for, and that alone. I suggest that anyone in the future needing fast random write IOPS is going to move those workloads to SSD, which is steadily increasing in capacity. And I suggest anyone building arrays with 10-20TB drives isn''t in need of fast random write IOPS. Whether this approach is valuable to anyone depends on whether the remaining attributes of RAID10, with the added URE protection, are worth the drive count. Obviously proponents of traditional parity arrays will not think so. Users of RAID10 may. Even if md never supports such a scheme, I bet we''ll see something similar to this in enterprise gear, where rebuilds need to be ''fast'' and performance degradation due to a downed drive is not acceptable. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 21/11/13 21:52, Piergiorgio Sartor wrote:> Hi David, > > On Thu, Nov 21, 2013 at 09:31:46PM +0100, David Brown wrote: > [...] >> If this can all be done to give the user an informed choice, then it >> sounds good. > > that would be my target. > To _offer_ more options to the (advanced) user. > It _must_ always be under user control. > >> One issue here is whether the check should be done with the filesystem >> mounted and in use, or only off-line. If it is off-line then it will >> mean a long down-time while the array is checked - but if it is online, >> then there is the risk of confusing the filesystem and caches by >> changing the data. > > Currently, "raid6check" can work with FS mounted. > I got the suggestion from Neil (of course). > It is possible to lock one stripe and check it. > This should be, at any given time, consistent > (that is, the parity should always match the data). > If an error is found, it is reported. > Again, the user can decide to fix it or not, > considering all the FS consequences and so on. >If you can lock stripes, and make sure any old data from that stripe is flushed from the caches (if you change it while locked), then that sounds ideal.>> Most disk errors /are/ detectable, and are reported by the underlying >> hardware - small surface errors are corrected by the disk''s own error >> checking and correcting mechanisms, and larger errors are usually >> detected. It is (or should be!) very rare that a read error goes >> undetected without there being a major problem with the disk controller. >> And if the error is detected, then the normal raid processing kicks in >> as there is no doubt about which block has problems. > > That''s clear. That case is an "erasure" (I think) > and it is perfectly in line with the usual operation. > I''m not trying to replace this mechanism. > >> If you can be /sure/ about which data block is incorrect, then I agree - >> but you can''t be /entirely/ sure. But I agree that you can make a good >> enough guess to recommend a fix to the user - as long as it is not >> automatic. > > One typical case is when many errors are > found, belonging to the same disk. > This case clearly shows the disk is to be > replaced or the interface checked... > But, again, the user is the master, not the > machine... :-)I don''t know what sort of interface you have for the user, but I guess that means you''ll have to collect a number of failures before showing them so that the user can see the correlation on disk number.> >> For most ECC schemes, you know that all your blocks are set >> synchronously - so any block that does not fit in, is an error. With >> raid, it could also be that a stripe is only partly written - you can > > Could it be? > I would consider this an error.It could occur as the result of a failure of some sort (kernel crash, power failure, temporary disk problem, etc.). More generally, md raid doesn''t have to be on local physical disks - maybe one of the "disks" is an iSCSI drive or something else over a network that could have failures or delays. I haven''t thought through all cases here - I am just throwing them out as possibilities that might cause trouble.> The stripe must always be consistent, there > should be a transactional mechanism to make > sure that, if read back, the data is always > matching the parity. > When I write "read back" I mean from whatever > the data is: physical disk or cache. > Otherwise, the check must run exclusively on > the array (no mounted FS, no other things > running on it). > >> have two different valid sets of data mixed to give an inconsistent >> stripe, without any good way of telling what consistent data is the best >> choice. >> >> Perhaps a checking tool can take advantage of a write-intent bitmap (if >> there is one) so that it knows if an inconsistent stripe is partly >> updated or the result of a disk error. > > Of course, this is an option, which should be > taken into consideration. > > Any improvement idea is welcome!!! > > bye, >-- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/21/2013 04:30 PM, Stan Hoeppner wrote:> > The rebuild time of a parity array normally has little to do with CPU > overhead.>Unless you have to fall back to table driven code. Anyway, this looks like a great concept. Now we just need to implement it ;) -hpa -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 22/11/13 01:30, Stan Hoeppner wrote:> I don''t like it either. It''s a compromise. But as RAID1/10 will soon > be unusable due to URE probability during rebuild, I think it''s a > relatively good compromise for some users, some workloads.An alternative is to move to 3-way raid1 mirrors rather than 2-way mirrors. Obviously you take another hit in disk space efficiency, but reads will be faster than you have extra redundancy. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, 21 Nov 2013 20:54:54 Adam Goryachev wrote:> On a pure storage server, the CPU would normally have nothing to do, > except a little interrupt handling, it is just shuffling bytes around. > Of course, if you need RAID7.5 then you probably have a dedicated > storage server, so I don''t see the problem with using the CPU to do all > the calculations.How would the pure storage server in question send data out to client systems? SMB? NFS? IMAP? All options will take some CPU power. There''s a common myth that people should use hardware RAID to save CPU time, that''s obviously wrong in the case of RAID-1 (just writing the data twice) and RAID-5 (XOR for parity). But as we get into the more complex forms of parity CPU performance could become an issue. Of course in terms of overall system performance the cost of doing a RAID rebuild in terms of disk seeks will probably exceed the CPU use. As an aside, are there plans to limit the RAID rebuild speed for BTRFS? On Thu, 21 Nov 2013 18:30:49 Stan Hoeppner wrote:> I suggest that anyone in the future needing fast random write IOPS is > going to move those workloads to SSD, which is steadily increasing in > capacity. And I suggest anyone building arrays with 10-20TB drives > isn''t in need of fast random write IOPS.Traditionally SCSI/SAS disks have tended to be a lot smaller than IDE/SATA disks. Now Dell has just started offering 2TB SAS disks while the largest SATA disks that they sell (In Australia on PowerEdge T110 servers at least) are also 2TB. Presumably RAID recovery time was one factor that made manufacturers not bother with making larger SCSI/SAS disks in the past. When 20TB disks become available a user could choose to just use the first 10TB of the disk. Even in the days when 40G disks were big some people would use less than the full capacity of the disk for performance. Will people really be storing data that is so important that triple parity is needed on 20TB disks but be unable to afford enough disks to use only the first 10TB of each disk? Is this a use-case that is worth coding for? -- My Main Blog http://etbe.coker.com.au/ My Documents Blog http://doc.coker.com.au/ -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi David, On 11/21/2013 3:07 AM, David Brown wrote:> On 21/11/13 02:28, Stan Hoeppner wrote:...>> WRT rebuild times, once drives hit 20TB we''re looking at 18 hours just >> to mirror a drive at full streaming bandwidth, assuming 300MB/s >> average--and that is probably being kind to the drive makers. With 6 or >> 8 of these drives, I''d guess a typical md/RAID6 rebuild will take at >> minimum 72 hours or more, probably over 100, and probably more yet for >> 3P. And with larger drive count arrays the rebuild times approach a >> week. Whose users can go a week with degraded performance? This is >> simply unreasonable, at best. I say it''s completely unacceptable. >> >> With these gargantuan drives coming soon, the probability of multiple >> UREs during rebuild are pretty high. Continuing to use ever more >> complex parity RAID schemes simply increases rebuild time further. The >> longer the rebuild, the more likely a subsequent drive failure due to >> heat buildup, vibration, etc. Thus, in our maniacal efforts to mitigate >> one failure mode we''re increasing the probability of another. TANSTAFL. >> Worse yet, RAID10 isn''t going to survive because UREs on a single drive >> are increasingly likely with these larger drives, and one URE during >> rebuild destroys the array.> I don''t think the chances of hitting an URE during rebuild is dependent > on the rebuild time - merely on the amount of data read during rebuild.Please read the above paragraph again, as you misread it the first time.> URE rates are "per byte read" rather than "per unit time", are they not?These are specified by the drive manufacturer, and they are per *bits* read, not "per byte read". Current consumer drives are typically rated at 1 URE in 10^14 bits read, enterprise are 1 in 10^15.> I think you are overestimating the rebuild times a bit, but there is noWhich part? A 20TB drive mirror taking 18 hours, or parity arrays taking many times longer than 18 hours?> arguing that rebuild on parity raids is a lot more work (for the cpu, > the IO system, and the disks) than for mirror raids.It''s not so much a matter of work or interface bandwidth, but a matter of serialization and rotational latency. ...> Shouldn''t we be talking about RAID 15 here, rather than RAID 51 ? I > interpret "RAID 15" to be like "RAID 10" - a raid5 set of raid1 mirrors, > while "RAID 51" would be a raid1 mirror of raid5 sets. I am certain > that you mean a raid5 set of raid1 pairs - I just think you''ve got the > name wrong.Now that you mention it, yes, RAID 15 would fit much better with convention. Not sure why I thought 51. So it''s RAID 15 from here.>> Potential Advantages: >> >> 1. Only +1 disk capacity overhead vs RAID 10, regardless of drive count > > +2 disks (the raid5 parity "disk" is a raid1 pair)One drive of each mirror is already gone. Make a RAID 5 of the remaining disks and you lose 1 disk. So you lose 1 additional disk vs RAID 10, not 2. As I stated previously, for RAID 15 you lose [1/2]+1 of your disks to redundancy. ...>> [1] The RAID1/5 code would need to be patched to properly handle a URE >> encountered by the RAID1 code during rebuild. There are surely other >> modifications and/or optimizations that would be needed. For large >> sequential reads, more deterministic read interleaving between mirror >> pairs would be a good candidate I think. IIUC the RAID1 driver does >> read interleaving on a per thread basis or some such, which I don''t >> believe is going to work for this "RAID 51" scenario, at least not for >> single streaming reads. If this can be done well, we double the read >> performance of RAID5, and thus we don''t completely "waste" all the extra >> disks vs big_parity schemes. >> >> This proposed "RAID level 51" should have drastically lower rebuild >> times vs traditional striped parity, should not suffer read/write >> performance degradation with most disk failure scenarios, and with a >> read interleaving optimization may have significantly greater streaming >> read throughput as well. >> >> This is far from a perfect solution and I am certainly not promoting it >> as such. But I think it does have some serious advantages over >> traditional striped parity schemes, and at minimum is worth discussion >> as a counterpoint of sorts. > > I don''t see that there needs to be any changes to the existing md code > to make raid15 work - it is merely a raid 5 made from a set of raid1 > pairs.The sole purpose of the parity layer of the proposed RAID 15 is to replace sectors lost due to UREs during rebuild. AFAIK the current RAID 5 and RAID 1 drivers have no code to support each other in this manner.> I can see that improved threading and interleaving could be a > benefit here - but that''s the case in general for md raid, and it is > something that the developers are already working on (I haven''t followed > the details, but the topic comes up regularly on the list here).What I''m talking about here is unrelated to the kernel thread starvation issue, which is write centric, unrelated to reads. What I''m suggesting is that it might be possible to improve the concurrency of reads from the mirror disks using some form of static or adaptive interleaving or similar. The purpose of this would be strictly to improve large single streaming read performance. If this could be achieved I do not know. One possibility may be to count consecutive LBA sectors requested by the filesystem stream and compare that to some value. For example, say we have an 18 disk RAID 15 array which gives us 8 spindles. With a default chunk of 512KB this gives us a stripe width of 4MB. So lets say we arbitrarily consider any single stream read larger than 4 stripes, 16MB, to be a large streaming read. So once our stream counter reaches 32,768 sectors we have the mirror code do alternating reads of 1,024 sectors, 512KB (chunk size), from each disk in the mirror. Theoretically, this could yield large streaming read performance double that of streaming write, and double that of the current RAID 1 read behavior on a per mirror basis. The trigger value could be statically defined at array creation time by a yet to be determined formula based on spindle count and chunk size, or it could be user configurable.> So as far as I can see, you''ve got raid15 support already - if that''s > what suits your needs, use it. Future improvements to the md code are > only needed to make it faster.You''re too hung up on names and not getting the point. Whether we call it RAID 15 or Blue Cheese, if it doesn''t have URE mitigation during rebuild, it''s worthless.> Of course, there is scope for making specific raid15 support in md along > the lines of the raid10 code - raid15,f2 would have the same speed > advantages over "normal" raid1+5 as raid10,f2 has over raid1+0. Whether > it is worth the effort implementing it is a different matter.Except the RAID10 driver suffers from the single write thread. RAID 0 over mirrors doesn''t have this problem. Which is why, along with other reasons, I proposed a possible RAID 15 driver using the RAID 5 and RAID 1 drivers as the base, as this won''t have the single write thread problem.> I can see plenty of reasons why raid15 might be a good idea, and even > raid16 for 5 disk redundancy, compared to multi-parity sets. However, > it costs a lot in disk space.... Of course it does, just as RAID 10 does. Parity users who currently shun RAID 10 for this reason will also shun this "RAID 15". That''s obvious. Potential users of RAID 15 are those who value the features of RAID 10 other than random write performance. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/21/2013 3:07 AM, David Brown wrote:> For example, with 20 disks at 1 TB each, you can have:All correct, and these are maximum redundancies. Maximum:> raid5 = 19TB, 1 disk redundancy > raid6 = 18TB, 2 disk redundancy > raid6.3 = 17TB, 3 disk redundancy > raid6.4 = 16TB, 4 disk redundancy > raid6.5 = 15TB, 5 disk redundancyThese are not fully correct, because only the minimums are stated. With any mirror based array one can lose half the disks as long as no two are in one mirror. The probability of a pair failing together is very low, and this probability decreases even further as the number of drives in the array increases. This is one of the many reasons RAID 10 has been so popular for so many years. Minimum:> raid10 = 10TB, 1 disk redundancy > raid15 = 8TB, 3 disk redundancy > raid16 = 6TB, 5 disk redundancyMaximum: RAID 10 = 10 disk redundancy RAID 15 = 11 disk redundancy RAID 16 = 12 disk redundancy Range: RAID 10 = 1-10 disk redundancy RAID 15 = 3-11 disk redundancy RAID 16 = 5-12 disk redundancy -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/21/2013 5:38 PM, John Williams wrote:> On Thu, Nov 21, 2013 at 2:57 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote: >> He wrote that article in late 2009. It seems pretty clear he wasn''t >> looking 10 years forward to 20TB drives, where the minimum mirror >> rebuild time will be ~18 hours, and parity rebuild will be much greater. > > Actually, it is completely obvious that he WAS looking ten years > ahead, seeing as several of his graphs have time scales going to > 2009+10 = 2019.Only one graph goes to 2019, the rest are 2010 or less. That being the case, his 2019 graph deals with projected reliability of single, double, and triple parity.> And he specifically mentions longer rebuild times as one of the > reasons why higher parity RAIDs are needed.Yes, he certainly does. But *only* in the context of the array surviving for the duration of a rebuild. He doesn''t state that he cares what the total duration is, he doesn''t guess what it might be, nor does he seem to care about the degraded performance before or during the rebuild. He is apparently of the mindset "more parity will save us, until we need more parity, until we need more parity, until we need more...". Following this path, parity will eventually eat more disks of capacity than RAID10 does today for average array counts, and the only reason for it being survival of ever increasing rebuild duration. This is precisely why I proposed "RAID 15". It gives you the single disk cloning rebuild speed of RAID 10. When parity hits 5P then RAID 15 becomes very competitive for smaller arrays. And since drives at that point will be 40-50TB each, even small arrays will need lots of protection against UREs and additional failures during massive rebuild times. Here I''d say RAID 15 will beat 5P hands down. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 22/11/13 09:13, Stan Hoeppner wrote:> Hi David, > > On 11/21/2013 3:07 AM, David Brown wrote: >> On 21/11/13 02:28, Stan Hoeppner wrote: > ... >>> WRT rebuild times, once drives hit 20TB we''re looking at 18 hours just >>> to mirror a drive at full streaming bandwidth, assuming 300MB/s >>> average--and that is probably being kind to the drive makers. With 6 or >>> 8 of these drives, I''d guess a typical md/RAID6 rebuild will take at >>> minimum 72 hours or more, probably over 100, and probably more yet for >>> 3P. And with larger drive count arrays the rebuild times approach a >>> week. Whose users can go a week with degraded performance? This is >>> simply unreasonable, at best. I say it''s completely unacceptable. >>> >>> With these gargantuan drives coming soon, the probability of multiple >>> UREs during rebuild are pretty high. Continuing to use ever more >>> complex parity RAID schemes simply increases rebuild time further. The >>> longer the rebuild, the more likely a subsequent drive failure due to >>> heat buildup, vibration, etc. Thus, in our maniacal efforts to mitigate >>> one failure mode we''re increasing the probability of another. TANSTAFL. >>> Worse yet, RAID10 isn''t going to survive because UREs on a single drive >>> are increasingly likely with these larger drives, and one URE during >>> rebuild destroys the array. > > >> I don''t think the chances of hitting an URE during rebuild is dependent >> on the rebuild time - merely on the amount of data read during rebuild. > > Please read the above paragraph again, as you misread it the first time.Yes, I thought you were saying that URE''s were more likely during a parity raid rebuild than during a mirror raid rebuild, because parity rebuilds take longer. They will be slightly more likely (due to more mechanical stress on the drives), but only slightly.> >> URE rates are "per byte read" rather than "per unit time", are they not? > > These are specified by the drive manufacturer, and they are per *bits* > read, not "per byte read". Current consumer drives are typically rated > at 1 URE in 10^14 bits read, enterprise are 1 in 10^15."Per bit" or "per byte" makes no difference to the principle. Just to get some numbers here, if we have a 20 TB drive (which don''t yet exist, AFAIK - 6 TB is the highest I have heard) with a URE rate of 1 in 10^14, that means an average of 1.6 errors per read of the whole disk. Assuming bit errors are independent (an unwarranted assumption, I know - but it makes the maths easier!), an URE of 1 in 10^14 gives a chance of 3.3 * 10^-10 of an error in a 4 KB sector - and a 83% chance of getting at least one incorrect sector read out of 20 TB. Even if enterprise disks have lower URE rates, I think it is reasonable to worry about URE''s during a raid1 rebuild! The probability of hitting URE''s on two disks at the same spot is, of course, tiny (given that you''ve got one URE, the chances of a URE in the same sector on another disk is 3.3 * 10^-10) - so two disk redundancy lets you have a disk failure and an URE safely. In theory, mirror raids are safer here because you only need to worry about a matching URE on /one/ disk. If you have a parity array with 60 disks, the chances of a matching URE on one of the other disks is 2 * 10^-8 - higher than for mirror raids, but still not a big concern. (Of course, you have more chance of a complete disk failure provoked by stresses during rebuilds, but that''s another failure mode.) What does all this mean? Single disk redundancy, like 2-way raid1 mirrors, is not going to be good enough for bigger disks unless the manufacturers can get their URE rates significantly lower. You will need an extra redundancy to be safe. That means raid6 is a minimum, or 3-way mirrors, or stacked raids like raid15. And if you want to cope with a disk failure, a second disk failure due to the stresses of rebuilding, /and/ and URE, then triple parity or raid15 is needed.> >> I think you are overestimating the rebuild times a bit, but there is no > > Which part? A 20TB drive mirror taking 18 hours, or parity arrays > taking many times longer than 18 hours?The 18 hours for a 20 TB mirror sounds right - but that it takes 9 times as long for a rebuild with a parity array sounds too much. But I don''t have any figures as evidence. And of course it varies depending on what else you are doing with the array at the time - parity array rebuilds will be affected much more by concurrent access to the array than mirrored arrays. It''s all a balance - if you want cheaper space but have less IO''s and can tolerate slower rebuilds, then parity arrays are good. If you want fast access then raid 15 looks better.> >> arguing that rebuild on parity raids is a lot more work (for the cpu, >> the IO system, and the disks) than for mirror raids. > > It''s not so much a matter of work or interface bandwidth, but a matter > of serialization and rotational latency. > > ... >> Shouldn''t we be talking about RAID 15 here, rather than RAID 51 ? I >> interpret "RAID 15" to be like "RAID 10" - a raid5 set of raid1 mirrors, >> while "RAID 51" would be a raid1 mirror of raid5 sets. I am certain >> that you mean a raid5 set of raid1 pairs - I just think you''ve got the >> name wrong. > > Now that you mention it, yes, RAID 15 would fit much better with > convention. Not sure why I thought 51. So it''s RAID 15 from here.Maybe you wanted to use the power of alien technology from Area 51 :-) But I''m glad we agree on the name.> >>> Potential Advantages: >>> >>> 1. Only +1 disk capacity overhead vs RAID 10, regardless of drive count >> >> +2 disks (the raid5 parity "disk" is a raid1 pair) > > One drive of each mirror is already gone. Make a RAID 5 of the > remaining disks and you lose 1 disk. So you lose 1 additional disk vs > RAID 10, not 2. As I stated previously, for RAID 15 you lose [1/2]+1 of > your disks to redundancy.Ah, you meant you lose a disk''s worth of capacity in comparison to a raid10 array with the same number of disks? I meant you have to add 2 disks to your raid10 array in order to keep the same capacity. Both are correct - it''s just a different way of looking at it. Just to be clear, to store data blocks D0, D1, D2, D3 on different raids you need: raid0: 4 disks D0, D1, D2, D3 raid10: 8 disks D0a, D0b, D1a, D1b, D2a, D2b, D3a, D3b raid5: 5 disks D0, D1, D2, D3, P raid15: 10 disks D0a, D0b, D1a, D1b, D2a, D2b, D3a, D3b, Pa, Pb> > ... >>> [1] The RAID1/5 code would need to be patched to properly handle a URE >>> encountered by the RAID1 code during rebuild. There are surely other >>> modifications and/or optimizations that would be needed. For large >>> sequential reads, more deterministic read interleaving between mirror >>> pairs would be a good candidate I think. IIUC the RAID1 driver does >>> read interleaving on a per thread basis or some such, which I don''t >>> believe is going to work for this "RAID 51" scenario, at least not for >>> single streaming reads. If this can be done well, we double the read >>> performance of RAID5, and thus we don''t completely "waste" all the extra >>> disks vs big_parity schemes. >>> >>> This proposed "RAID level 51" should have drastically lower rebuild >>> times vs traditional striped parity, should not suffer read/write >>> performance degradation with most disk failure scenarios, and with a >>> read interleaving optimization may have significantly greater streaming >>> read throughput as well. >>> >>> This is far from a perfect solution and I am certainly not promoting it >>> as such. But I think it does have some serious advantages over >>> traditional striped parity schemes, and at minimum is worth discussion >>> as a counterpoint of sorts. >> >> I don''t see that there needs to be any changes to the existing md code >> to make raid15 work - it is merely a raid 5 made from a set of raid1 >> pairs. > > The sole purpose of the parity layer of the proposed RAID 15 is to > replace sectors lost due to UREs during rebuild. AFAIK the current RAID > 5 and RAID 1 drivers have no code to support each other in this manner.Now I''ve figure out what you are thinking about - if the raid1 rebuild fails on a stripe due to an URE, then rather than just marking that stripe as bad, it should ask the higher raid level (the raid5 here) for the data. I don''t know if there is any such system in place at the moment in the kernel code - maybe one of the md code experts here can tell us. It should be reasonably easy to implement, I think - what is needed is for a failure on the lower level raid to trigger a scrub on the stripe at the higher level raid. (Of course, the problem will solve itself next time you do a scrub on the upper raid anyway, but it would be best to fix it quickly.) One other optimisation that could be nice here when rebuilding one of the mirror pairs is to mark the pair "write-mostly", and possibly even "write-behind". These flags are currently only valid for raid1 (AFAIK), and can only be set when building an array. The idea here is that you can have a mirror between a local fast drive and a slower drive or a networked drive - the slow drive will only be used for writes unless there is a failure on the faster drives, and writes can be buffered (with "write-behind") if needed. If a rebuilding mirror pair in raid 15 could be temporarily marked as "write-mostly" and perhaps "write-behind", then it would be free to dedicate full bandwidth to the rebuild. Any reads from that pair would be re-created from the parities on the other drives.> >> I can see that improved threading and interleaving could be a >> benefit here - but that''s the case in general for md raid, and it is >> something that the developers are already working on (I haven''t followed >> the details, but the topic comes up regularly on the list here). > > What I''m talking about here is unrelated to the kernel thread starvation > issue, which is write centric, unrelated to reads. > > What I''m suggesting is that it might be possible to improve the > concurrency of reads from the mirror disks using some form of static or > adaptive interleaving or similar. The purpose of this would be strictly > to improve large single streaming read performance. If this could be > achieved I do not know. > > One possibility may be to count consecutive LBA sectors requested by the > filesystem stream and compare that to some value. For example, say we > have an 18 disk RAID 15 array which gives us 8 spindles. With a default > chunk of 512KB this gives us a stripe width of 4MB. So lets say we > arbitrarily consider any single stream read larger than 4 stripes, 16MB, > to be a large streaming read. So once our stream counter reaches 32,768 > sectors we have the mirror code do alternating reads of 1,024 sectors, > 512KB (chunk size), from each disk in the mirror. >I see what you mean, and yes, I can see how that could speed up large accesses. But I think the same idea applies to a normal raid1 mirror - a streamed read from it could use interleaved reads from the two halves to speed up the read. And if it worked for raid1, then it would automatically work for raid 15. I think :-) Of course, you could always use raid10,f2 or raid10,o2 under raid5 to get faster read performance (at the cost of slower writes - just like normal raid10,f2 vs. raid1 comparisons). Then on your 18-disk "raid105" array your streamed reads would use 16 spindles.> Theoretically, this could yield large streaming read performance double > that of streaming write, and double that of the current RAID 1 read > behavior on a per mirror basis. The trigger value could be statically > defined at array creation time by a yet to be determined formula based > on spindle count and chunk size, or it could be user configurable. > >> So as far as I can see, you''ve got raid15 support already - if that''s >> what suits your needs, use it. Future improvements to the md code are >> only needed to make it faster. > > You''re too hung up on names and not getting the point. Whether we call > it RAID 15 or Blue Cheese, if it doesn''t have URE mitigation during > rebuild, it''s worthless.As noted above, I think the current system would work but you need a scrub at the raid5 level after the raid1 rebuild - and (now that I see what you are getting at) I agree that this could be done much better with extra kernel support.> >> Of course, there is scope for making specific raid15 support in md along >> the lines of the raid10 code - raid15,f2 would have the same speed >> advantages over "normal" raid1+5 as raid10,f2 has over raid1+0. Whether >> it is worth the effort implementing it is a different matter. > > Except the RAID10 driver suffers from the single write thread. RAID 0 > over mirrors doesn''t have this problem. Which is why, along with other > reasons, I proposed a possible RAID 15 driver using the RAID 5 and RAID > 1 drivers as the base, as this won''t have the single write thread problem.Improving the threading of raid10 writes should be perfectly possible technically - the only problem is someone having the time to do it. We are just looking at different ideas here, to gauge the pros and cons of alternative raid structures. Ideally we could have /all/ these ideas implemented, but developer and tester time is the limitation rather than the technical solution (since it looks like the technical problems of multi-parity raid have been solved).> >> I can see plenty of reasons why raid15 might be a good idea, and even >> raid16 for 5 disk redundancy, compared to multi-parity sets. However, >> it costs a lot in disk space. > ... > > Of course it does, just as RAID 10 does. Parity users who currently > shun RAID 10 for this reason will also shun this "RAID 15". That''s > obvious. Potential users of RAID 15 are those who value the features of > RAID 10 other than random write performance. >Yes, both solutions would be useful. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 22/11/13 09:38, Stan Hoeppner wrote:> On 11/21/2013 3:07 AM, David Brown wrote: > >> For example, with 20 disks at 1 TB each, you can have: > > All correct, and these are maximum redundancies. > > Maximum: > >> raid5 = 19TB, 1 disk redundancy >> raid6 = 18TB, 2 disk redundancy >> raid6.3 = 17TB, 3 disk redundancy >> raid6.4 = 16TB, 4 disk redundancy >> raid6.5 = 15TB, 5 disk redundancy > > > These are not fully correct, because only the minimums are stated. With > any mirror based array one can lose half the disks as long as no two are > in one mirror. The probability of a pair failing together is very low, > and this probability decreases even further as the number of drives in > the array increases. This is one of the many reasons RAID 10 has been > so popular for so many years. > > Minimum: > >> raid10 = 10TB, 1 disk redundancy >> raid15 = 8TB, 3 disk redundancy >> raid16 = 6TB, 5 disk redundancy > > Maximum: > > RAID 10 = 10 disk redundancy > RAID 15 = 11 disk redundancy12 disks maximum (you have 8 with data, the rest are mirrors, parity, or mirrors of parity).> RAID 16 = 12 disk redundancy14 disks maximum (you have 6 with data, the rest are mirrors, parity, or mirrors of parity).> > Range: > > RAID 10 = 1-10 disk redundancy > RAID 15 = 3-11 disk redundancy > RAID 16 = 5-12 disk redundancy > >Yes, I know these are the minimum redundancies. But that''s a vital figure for reliability (even if the range is important for statistical averages). When one disk in a raid10 array fails, your main concern is about failures or URE''s in the other half of the pair - it doesn''t help to know that another nine disks can "safely" fail too. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, Nov 22, 2013 at 1:35 AM, Stan Hoeppner <stan@hardwarefreak.com> wrote:> Only one graph goes to 2019, the rest are 2010 or less. That being the > case, his 2019 graph deals with projected reliability of single, double, > and triple parity.The whole article goes to 2019 (or longer). He shows current trends and discusses where they are going in the future. The whole point of the article is looking ahead into the future.> Following this path, parity will eventually eat more disks of capacity > than RAID10 does today for average array counts, and the only reason for > it being survival of ever increasing rebuild duration.No, that is not what the article finds. In the near future (about 10 years), triple-parity will suffice. Beyond that, perhaps quad-parity will be required, but predicting that far ahead is usually worthless in the computer industry.> When parity hits 5P then RAID 15 > becomes very competitive for smaller arrays. And since drives at that > point will be 40-50TB each, even small arrays will need lots of > protection against UREs and additional failures during massive rebuild > times. Here I''d say RAID 15 will beat 5P hands down.I''ll take triple- or 4-parity every time over the disk-wasting and less reliable RAID 15. There is no need for 5-parity in the near future. I see no advantage of RAID 15, and several disadvantages. -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/22/2013 2:13 AM, Stan Hoeppner wrote:> Hi David, > > On 11/21/2013 3:07 AM, David Brown wrote:...>> I don''t see that there needs to be any changes to the existing md code >> to make raid15 work - it is merely a raid 5 made from a set of raid1 >> pairs. > > The sole purpose of the parity layer of the proposed RAID 15 is to > replace sectors lost due to UREs during rebuild. AFAIK the current RAID > 5 and RAID 1 drivers have no code to support each other in this manner.Minor self correction here-- obviously this isn''t the ''sole'' purpose of the parity layer. It also allows us to recover from losing an entire mirror, which is a big upshot of the proposed RAID 15. Thinking this through a little further, more code modification would be needed for this scenario. In the event of a double drive failure in one mirror, the RAID 1 code will need to be modified in such a way as to allow the RAID 5 code to rebuild the first replacement disk, because the RAID 1 device is still in a failed state. Once this rebuild is complete, the RAID 1 code will need to switch the state to degraded, and then do its standard rebuild routine for the 2nd replacement drive. Or, with some (likely major) hacking it should be possible to rebuild both drives simultaneously for no loss of throughput or additional elapsed time on the RAID 5 rebuild. In the 20TB drive case, this would shave 18 hours off the total rebuild operation elapsed time. With current 4TB drives it would still save 6.5 hours. Losing both drives in one mirror set of a striped array is rare, but given the rebuild time saved it may be worth investigating during any development of this RAID 15 idea. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, Nov 22, 2013 at 12:13 AM, Stan Hoeppner <stan@hardwarefreak.com> wrote:> Hi David, > > On 11/21/2013 3:07 AM, David Brown wrote:<SNIP>>> Shouldn''t we be talking about RAID 15 here, rather than RAID 51 ? I >> interpret "RAID 15" to be like "RAID 10" - a raid5 set of raid1 mirrors, >> while "RAID 51" would be a raid1 mirror of raid5 sets. I am certain >> that you mean a raid5 set of raid1 pairs - I just think you''ve got the >> name wrong. > > Now that you mention it, yes, RAID 15 would fit much better with > convention. Not sure why I thought 51. So it''s RAID 15 from here.<SNIP> For us casual readers & RAID users could you clarify RAID15? Would that be a bunch of RAID1''s grouped together in what appears to be a RAID5 to the system? Thanks, Mark -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Mark Knecht posted on Fri, 22 Nov 2013 08:50:32 -0800 as excerpted:> On Fri, Nov 22, 2013 at 12:13 AM, Stan Hoeppner <stan@hardwarefreak.com> > wrote: >> Now that you mention it, yes, RAID 15 would fit much better with >> convention. Not sure why I thought 51. So it''s RAID 15 from here. > <SNIP> > > For us casual readers & RAID users could you clarify RAID15? Would that > be a bunch of RAID1''s grouped together in what appears to be a RAID5 to > the system?Simplest definition, yes. Admittedly part of this discussion is beyond me (as another casual reader with some raid experience, reading here via the btrfs list as that''s my current interest), but I''m following enough of it to find it interesting, for SURE! =:^) And perhaps my explanation of the basics will let the real experts continue the debate at their higher level... At a concept level, because md/raid, etc (I''ll use mdraid as my example from here, but but there''s dm-raid, hardware raid, etc; additionally, I''ll omit the ALL CAPS RAID convention and use lowercase), devices are presented as normal block devices, RAID levels (among other things, LVM2, etc) are stackable. So it''s possible to, for instance, create a raid0 on top of a bunch of raid1s, or the reverse, a raid1 on top of a bunch of raid0s, either with the base level being hardware based and the software creating a raid level direct on the hardware raid, or with both/all levels in software. Then we get into naming. AFAIK the earliest convention was using the plus syntax, raid1+0, raid0+1, with the left-most number being the lowest, closest to hardware level, either the hardware level or closest to the individual hardware devices, so raid1+0 is implemented as striped raid (raid0) over top of mirrored raid (raid1), with raid0+1 the reverse, a mirror over stripes. That quickly evolved into omitting the +, thus raid10 and raid01. (Tho 01 has the leading zero problem with some people trying to omit it, and raid1 isn''t the same thing AT ALL as raid01! Between that and the fact that raid01 is less common than raid10 for technical reasons as noted below, you seldom see raid01 specified; it usually keeps the + and appears as raid0+1). Also, less commonly seen but as more levels were stacked (raid105, etc), sometimes the + is still used to separate the hardware raid levels from software. In this usage, raid105 would probably be an all software implementation, while raid1+05 would be raid1 in hardware, with software raid0 and raid5 stacked on top, in that order, and raid10+5 would be hardware raid10, with software raid5 on top. Note that while raid10, aka raid1+0, should have similar non-degraded performance to raid0+1, there''s a BIG difference when recovering from degraded. A smart raid10 implementation (or a raid1+0 with hardware raid1) can rebuild a failed drive "locally", that is, purely at the raid1 level, using just the data on its raid1 mirror(s). That means only a single device has to be read from in ordered to write the data to the rebuilding device. Raid0+1, by contrast, fails the entire raid0 level at once, thus requiring reading from an unfailed entire raid1 (higher) level mirror set while writing out an entire new raid0 set!! So while normal operation state is similar between raid10/raid1+0 and raid0+1, the recovery characteristics are **MUCH** different, with raid10 being markedly better than raid0+1. As a result, raid0+1 doesn''t tend to be used that often in practice, while raid10 (aka raid1+0) has become quite common, particularly so as its performance is quite high, only exceeded by raid0, but with redundancy and recovery characteristics that are good to very good, as well. Its biggest negative at the low end is the number of devices required, normally a minimum of four (but see the Linux mdraid10 discussion below), a striped pair of mirrored pairs. This 1+0/0+1 distinction confused me as an early raid user for quite some time even after I knew the technical difference, as I kept trying to reverse them in my head, and I guess it confuses a lot of people. For some reason, my intuitive read of raid10 was the reverse of convention -- intuitively I /wanted/ to interpret it as a raid1 on top of raid0 instead of the raid0 on top of raid1 it is by convention, and even after I understood that there WAS a difference and in principle knew why and how, for years I actually had to look up the difference each time it came up, if it made a difference to the discussion, because I /wanted/ to read it backward, or more accurately, I thought the convention had it backward to the interpretation that made most sense to me. It is only recently that I came to see it the other way, and even still, I have to pause and think every time I see it, to ensure I''m not again reversing things. Which is the distinction that came up in the above discussion as well, only with raid5 and raid1 instead of raid0 and raid1. Apparently I''m not the only one to get things reversed! But yes, conceptually, raid15 is a raid5 layer on top of raid1, aka raid1 +5, while raid51 would be a raid1 layer on top of raid5, aka raid5+1. For the same recovery-time reasons noted above with raid0+1 vs. raid1+0/ raid10, having raid1 at the local/hardware layer should be preferable. With the basic concepts covered, the next level up is understanding that the Linux md/raid10 implementation, while BASED on the raid10 concept above, has some quite interesting extensions. Implementing it as a single software raid10 level instead of separate raid0 over raid1 allows some interesting optimizations and additional flexibility. Among other things, it no longer requires a minimum four devices (a raid0 pair of raid1 pairs) as separate raid0 over raid1 would. There''s quite a bit of additional flexibility in layout. A detailed discussion is out of scope here, but googling raid10 on wikipedia is a good start, and the page it gives you actually discusses various other nested raid levels as well. From there, follow the links to non-standard raid levels, and to the Linux mdraid implementation discussion, including the concepts of "near", "far", and "offset" layouts. https://en.wikipedia.org/wiki/Raid10 https://en.wikipedia.org/wiki/Non-standard_RAID_levels https://en.wikipedia.org/wiki/Non-standard_RAID_levels#Linux_MD_RAID_10 But the discussion here is well beyond that, out toward further implementation detail and optimization. One of the problems that has been creeping up on us is the fact that as shear drive sizes increase, the possibility of undetected/uncorrected physical device errors goes up faster than the technology gets better at reducing them. For "simple" parity RAID solutions such as raid5, this is a rather big problem, because at some point, the chances of error during recovery scuttling the recovery entirely simply get too large to practically deal with, with recovery time (and thus time to recovery failure and try again) similarly increasing toward the days and weeks point. If recovery''s going to take days, only for it to fail due to physical device error forcing another try... So the discussion is how to mitigate the problem. Multi-way-parity is of course the primary discussion in this thread, allowing detection and recovery of single-sector physical device errors via N-way-parity. But an integrated raid15 solution similar to mdraid''s current raid10, is another possibility, effectively using the raid1 mirror level to mitigate sector-level physical device errors, while using the higher raid5 level to detect them and trigger a re-mirror at the raid1 level below it. But the only way that can work is if the two conceptually separate raid levels are integrated at the implementation level, so the raid5 level parity error detection can tell the raid1 level which of its mirrors is bad and force a remirroring from the good one(s) to the bad one. -- Duncan - List replies preferred. No HTML msgs. "Every nonfree program has a lord, a master -- and if you use the program, he is your master." Richard Stallman -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi David, On Fri, Nov 22, 2013 at 01:32:09AM +0100, David Brown wrote:> > One typical case is when many errors are > > found, belonging to the same disk. > > This case clearly shows the disk is to be > > replaced or the interface checked... > > But, again, the user is the master, not the > > machine... :-) > > I don''t know what sort of interface you have for the user, but I guess > that means you''ll have to collect a number of failures before showing > them so that the user can see the correlation on disk number.as usual in Unix, one software will collect data to a file, an other one will analyze that file. Originally, one idea was even to check at stripe level how many errors (and where) are present. From that some statistics will be presented to the user. This would be integrated in the check tool, of course.> >> For most ECC schemes, you know that all your blocks are set > >> synchronously - so any block that does not fit in, is an error. With > >> raid, it could also be that a stripe is only partly written - you can > > > > Could it be? > > I would consider this an error. > > It could occur as the result of a failure of some sort (kernel crash, > power failure, temporary disk problem, etc.). More generally, md raid > doesn''t have to be on local physical disks - maybe one of the "disks" is > an iSCSI drive or something else over a network that could have failures > or delays. I haven''t thought through all cases here - I am just > throwing them out as possibilities that might cause trouble.OK, I misunderstood you, I was thinking during normal operation... Again, the check can find that issue, it will tell that it cannot find where the problem is. But it will tell where. Possibly, an other tool can check the FS at that position. bye, -- piergiorgio -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/22/2013 9:01 AM, John Williams wrote: <snip>> I see no advantage of RAID 15, and several disadvantages.Of course not, just as I sated previously. On 11/22/2013 2:13 AM, Stan Hoeppner wrote:> Parity users who currently shun RAID 10 for this reason will also > shun this "RAID 15".With that I''ll thank you for your input from the pure parity perspective, and end our discussion. Any further exchange would be pointless. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 22 Nov 2013 10:07:09 -0600 Stan Hoeppner <stan@hardwarefreak.com> wrote:> On 11/22/2013 2:13 AM, Stan Hoeppner wrote: > > Hi David, > > > > On 11/21/2013 3:07 AM, David Brown wrote: > ... > >> I don''t see that there needs to be any changes to the existing md code > >> to make raid15 work - it is merely a raid 5 made from a set of raid1 > >> pairs. > > > > The sole purpose of the parity layer of the proposed RAID 15 is to > > replace sectors lost due to UREs during rebuild. AFAIK the current RAID > > 5 and RAID 1 drivers have no code to support each other in this manner. > > Minor self correction here-- obviously this isn''t the ''sole'' purpose of > the parity layer. It also allows us to recover from losing an entire > mirror, which is a big upshot of the proposed RAID 15. Thinking this > through a little further, more code modification would be needed for > this scenario. > > In the event of a double drive failure in one mirror, the RAID 1 code > will need to be modified in such a way as to allow the RAID 5 code to > rebuild the first replacement disk, because the RAID 1 device is still > in a failed state. Once this rebuild is complete, the RAID 1 code will > need to switch the state to degraded, and then do its standard rebuild > routine for the 2nd replacement drive. > > Or, with some (likely major) hacking it should be possible to rebuild > both drives simultaneously for no loss of throughput or additional > elapsed time on the RAID 5 rebuild.Nah, that would be minor hacking. Just recreate the RAID1 in a state that is not-insync, but with automatic-resync disabled. Then as continuous writes arrive, move the "recovery_cp" variable forward towards the end of the array. When it reaches the end we can safely mark the whole array as ''in-sync'' and forget about diabling auto-resync. NeilBrown In the 20TB drive case, this would> shave 18 hours off the total rebuild operation elapsed time. With > current 4TB drives it would still save 6.5 hours. Losing both drives in > one mirror set of a striped array is rare, but given the rebuild time > saved it may be worth investigating during any development of this RAID > 15 idea. >
On Thu, 21 Nov 2013 16:57:48 -0600 Stan Hoeppner <stan@hardwarefreak.com> wrote:> On 11/21/2013 1:05 AM, John Williams wrote: > > On Wed, Nov 20, 2013 at 10:52 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote: > >> On 11/20/2013 8:46 PM, John Williams wrote: > >>> For myself or any machines I managed for work that do not need high > >>> IOPS, I would definitely choose triple- or quad-parity over RAID 51 or > >>> similar schemes with arrays of 16 - 32 drives. > >> > >> You must see a week long rebuild as acceptable... > > > > It would not be a problem if it did take that long, since I would have > > extra parity units as backup in case of a failure during a rebuild. > > > > But of course it would not take that long. Take, for example, a 24 x > > 3TB triple-parity array (21+3) that has had two drive failures > > (perhaps the rebuild started with one failure, but there was soon > > another failure). I would expect the rebuild to take about a day. > > You''re looking at today. We''re discussing tomorrow''s needs. Today''s > 6TB 3.5" drives have sustained average throughput of ~175MB/s. > Tomorrow''s 20TB drives will be lucky to do 300MB/s. As I said > previously, at that rate a straight disk-disk copy of a 20TB drive takes > 18.6 hours. This is what you get with RAID1/10/51. In the real world, > rebuilding a failed drive in a 3P array of say 8 of these disks will > likely take at least 3 times as long, 2 days 6 hours minimum, probably > more. This may be perfectly acceptable to some, but probably not to all.Could you explain your logic here? Why do you think rebuilding parity will take 3 times as long as rebuilding a copy? Can you measure that sort of difference today? Presumably when we have 20TB drives we will also have more cores and quite possibly dedicated co-processors which will make the CPU load less significant. NeilBrown
On 11/22/2013 5:07 PM, NeilBrown wrote:> On Thu, 21 Nov 2013 16:57:48 -0600 Stan Hoeppner <stan@hardwarefreak.com> > wrote: > >> On 11/21/2013 1:05 AM, John Williams wrote: >>> On Wed, Nov 20, 2013 at 10:52 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote: >>>> On 11/20/2013 8:46 PM, John Williams wrote: >>>>> For myself or any machines I managed for work that do not need high >>>>> IOPS, I would definitely choose triple- or quad-parity over RAID 51 or >>>>> similar schemes with arrays of 16 - 32 drives. >>>> >>>> You must see a week long rebuild as acceptable... >>> >>> It would not be a problem if it did take that long, since I would have >>> extra parity units as backup in case of a failure during a rebuild. >>> >>> But of course it would not take that long. Take, for example, a 24 x >>> 3TB triple-parity array (21+3) that has had two drive failures >>> (perhaps the rebuild started with one failure, but there was soon >>> another failure). I would expect the rebuild to take about a day. >> >> You''re looking at today. We''re discussing tomorrow''s needs. Today''s >> 6TB 3.5" drives have sustained average throughput of ~175MB/s. >> Tomorrow''s 20TB drives will be lucky to do 300MB/s. As I said >> previously, at that rate a straight disk-disk copy of a 20TB drive takes >> 18.6 hours. This is what you get with RAID1/10/51. In the real world, >> rebuilding a failed drive in a 3P array of say 8 of these disks will >> likely take at least 3 times as long, 2 days 6 hours minimum, probably >> more. This may be perfectly acceptable to some, but probably not to all. > > Could you explain your logic here? Why do you think rebuilding parity > will take 3 times as long as rebuilding a copy? Can you measure that sort of > difference today?I''ve not performed head-to-head timed rebuild tests of mirror vs parity RAIDs. I''m making the elapsed guess for parity RAIDs based on posts here over the past ~3 years, in which many users reported 16-24+ hour rebuild times for their fairly wide (12-16 1-2TB drive) RAID6 arrays. This is likely due to their chosen rebuild priority and concurrent user load during rebuild. Since this seems to be the norm, instead of giving 100% to the rebuild, I thought it prudent to take this into account, instead of the theoretical minimum rebuild time.> Presumably when we have 20TB drives we will also have more cores and quite > possibly dedicated co-processors which will make the CPU load less > significant.But (when) will we have the code to fully take advantage of these? It''s nearly 2014 and we still don''t have a working threaded write model for levels 5/6/10, though maybe soon. Multi-core mainstream x86 CPUs have been around for 8 years now, SMP and ccNUMA systems even longer. So the need has been there for a while. I''m strictly making an observation (possibly not fully accurate) here. I am not casting stones. I''m not a programmer and am thus unable to contribute code, only ideas and troubleshooting assistance for fellow users. Ergo I have no right/standing to complain about the rate of feature progress. I know that everyone hacking md is making the most of the time they have available. So again, not a complaint, just an observation. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 22 Nov 2013 21:46:50 -0600 Stan Hoeppner <stan@hardwarefreak.com> wrote:> On 11/22/2013 5:07 PM, NeilBrown wrote: > > On Thu, 21 Nov 2013 16:57:48 -0600 Stan Hoeppner <stan@hardwarefreak.com> > > wrote: > > > >> On 11/21/2013 1:05 AM, John Williams wrote: > >>> On Wed, Nov 20, 2013 at 10:52 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote: > >>>> On 11/20/2013 8:46 PM, John Williams wrote: > >>>>> For myself or any machines I managed for work that do not need high > >>>>> IOPS, I would definitely choose triple- or quad-parity over RAID 51 or > >>>>> similar schemes with arrays of 16 - 32 drives. > >>>> > >>>> You must see a week long rebuild as acceptable... > >>> > >>> It would not be a problem if it did take that long, since I would have > >>> extra parity units as backup in case of a failure during a rebuild. > >>> > >>> But of course it would not take that long. Take, for example, a 24 x > >>> 3TB triple-parity array (21+3) that has had two drive failures > >>> (perhaps the rebuild started with one failure, but there was soon > >>> another failure). I would expect the rebuild to take about a day. > >> > >> You''re looking at today. We''re discussing tomorrow''s needs. Today''s > >> 6TB 3.5" drives have sustained average throughput of ~175MB/s. > >> Tomorrow''s 20TB drives will be lucky to do 300MB/s. As I said > >> previously, at that rate a straight disk-disk copy of a 20TB drive takes > >> 18.6 hours. This is what you get with RAID1/10/51. In the real world, > >> rebuilding a failed drive in a 3P array of say 8 of these disks will > >> likely take at least 3 times as long, 2 days 6 hours minimum, probably > >> more. This may be perfectly acceptable to some, but probably not to all. > > > > Could you explain your logic here? Why do you think rebuilding parity > > will take 3 times as long as rebuilding a copy? Can you measure that sort of > > difference today? > > I''ve not performed head-to-head timed rebuild tests of mirror vs parity > RAIDs. I''m making the elapsed guess for parity RAIDs based on posts > here over the past ~3 years, in which many users reported 16-24+ hour > rebuild times for their fairly wide (12-16 1-2TB drive) RAID6 arrays.I guess with that many drives you could hit PCI bus throughput limits. A 16-lane PCIe 4.0 could just about give 100MB/s to each of 16 devices. So you would really need top-end hardware to keep all of 16 drives busy in a recovery. So yes: rebuilding a drive in a 16-drive RAID6+ would be slower than in e.g. a 20 drive RAID10.> > This is likely due to their chosen rebuild priority and concurrent user > load during rebuild. Since this seems to be the norm, instead of giving > 100% to the rebuild, I thought it prudent to take this into account, > instead of the theoretical minimum rebuild time. > > > Presumably when we have 20TB drives we will also have more cores and quite > > possibly dedicated co-processors which will make the CPU load less > > significant. > > But (when) will we have the code to fully take advantage of these? It''s > nearly 2014 and we still don''t have a working threaded write model for > levels 5/6/10, though maybe soon. Multi-core mainstream x86 CPUs have > been around for 8 years now, SMP and ccNUMA systems even longer. So the > need has been there for a while.I think we might have that multi-threading now - not sure exactly what is enabled by default though. I think it requires more than "need" - it requires "demand". i.e. people repeatedly expressing the need. We certainly have had that for a while, but not a very long while> > I''m strictly making an observation (possibly not fully accurate) here. > I am not casting stones. I''m not a programmer and am thus unable to > contribute code, only ideas and troubleshooting assistance for fellow > users. Ergo I have no right/standing to complain about the rate of > feature progress. I know that everyone hacking md is making the most of > the time they have available. So again, not a complaint, just an > observation.Understood - and thanks for your observation. NeilBrown
On Fri, Nov 22, 2013 at 9:04 PM, NeilBrown <neilb@suse.de> wrote:> I guess with that many drives you could hit PCI bus throughput limits. > > A 16-lane PCIe 4.0 could just about give 100MB/s to each of 16 devices. So > you would really need top-end hardware to keep all of 16 drives busy in a > recovery. > So yes: rebuilding a drive in a 16-drive RAID6+ would be slower than in e.g. > a 20 drive RAID10.Not really. A single 8x PCIe 2.0 card has 8 x 500MB/s = 4000MB/s of potential bandwidth. That would be 250MB/s per drive for 16 drives. But quite a few people running software RAID with many drives have multiple PCIe cards. For example, in one machine I have three IBM M1015 cards (which I got for $75/ea) that are 8x PCIe 2.0. That comes to 3 x 500MB/s x 8 = 12GB/s of IO bandwidth. Also, your math is wrong. PCIe 3.0 is 985 MB/s per lane. If we assume PCIe 4.0 would double that, we would have 1970MB/s per lane. So one lane of the hypothetical PCIe 4.0 would have enough IO bandwidth to give about 120MB/s to each of 16 drives. A single 8x PCIe 4.0 card would have 8 times that capability which is more than 15GB/s. Even a single 8x PCIe 3.0 card has potentially over 7GB/s of bandwidth. Bottom line is that IO bandwidth is not a problem for a system with prudently chosen hardware. More likely is that you would be CPU limited (rather than bus limited) in a high-parity rebuild where more than one drive failed. But even that is not likely to be too bad, since Andrea''s single-threaded recovery code can recover two drives at nearly 1GB/s on one of my machines. I think the code could probably be threaded to achieve a multiple of that running on multiple cores. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 22 Nov 2013 21:34:41 -0800 John Williams <jwilliams4200@gmail.com> wrote:> On Fri, Nov 22, 2013 at 9:04 PM, NeilBrown <neilb@suse.de> wrote: > > > I guess with that many drives you could hit PCI bus throughput limits. > > > > A 16-lane PCIe 4.0 could just about give 100MB/s to each of 16 devices. So > > you would really need top-end hardware to keep all of 16 drives busy in a > > recovery. > > So yes: rebuilding a drive in a 16-drive RAID6+ would be slower than in e.g. > > a 20 drive RAID10. > > Not really. A single 8x PCIe 2.0 card has 8 x 500MB/s = 4000MB/s of > potential bandwidth. That would be 250MB/s per drive for 16 drives. > > But quite a few people running software RAID with many drives have > multiple PCIe cards. For example, in one machine I have three IBM > M1015 cards (which I got for $75/ea) that are 8x PCIe 2.0. That comes > to 3 x 500MB/s x 8 = 12GB/s of IO bandwidth. > > Also, your math is wrong. PCIe 3.0 is 985 MB/s per lane. If we assume > PCIe 4.0 would double that, we would have 1970MB/s per lane. So one > lane of the hypothetical PCIe 4.0 would have enough IO bandwidth to > give about 120MB/s to each of 16 drives. A single 8x PCIe 4.0 card > would have 8 times that capability which is more than 15GB/s.It wasn''t my math, it was my reading :-( 16-lane PCIe 4.0 is 31 GB/sec so 2GB/sec per drive. I was reading the "1-lane" number...> > Even a single 8x PCIe 3.0 card has potentially over 7GB/s of bandwidth. > > Bottom line is that IO bandwidth is not a problem for a system with > prudently chosen hardware. > > More likely is that you would be CPU limited (rather than bus limited) > in a high-parity rebuild where more than one drive failed. But even > that is not likely to be too bad, since Andrea''s single-threaded > recovery code can recover two drives at nearly 1GB/s on one of my > machines. I think the code could probably be threaded to achieve a > multiple of that running on multiple cores.Indeed. It seems likely that with modern hardware, the linear write speed would be the limiting factor for spinning-rust drives. For SSDs the limit might end up being somewhere else ... Thanks, NeilBrown
Hi Piergiorgio,> How about par2? How does this work?I checked the matrix they use, and sometimes it contains some singular square submatrix. It seems that in GF(2^16) these cases are just less common. Maybe they were just unnoticed. Anyway, this seems to be an already known problem for PAR2, with an hypothetical PAR3 fixing it: http://sourceforge.net/p/parchive/discussion/96282/thread/d3c6597b/ Ciao, Andrea -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 22/11/13 23:59, NeilBrown wrote:> On Fri, 22 Nov 2013 10:07:09 -0600 Stan Hoeppner <stan@hardwarefreak.com> > wrote: ><snip>>> In the event of a double drive failure in one mirror, the RAID 1 code >> will need to be modified in such a way as to allow the RAID 5 code to >> rebuild the first replacement disk, because the RAID 1 device is still >> in a failed state. Once this rebuild is complete, the RAID 1 code will >> need to switch the state to degraded, and then do its standard rebuild >> routine for the 2nd replacement drive. >> >> Or, with some (likely major) hacking it should be possible to rebuild >> both drives simultaneously for no loss of throughput or additional >> elapsed time on the RAID 5 rebuild. > > Nah, that would be minor hacking. Just recreate the RAID1 in a state that is > not-insync, but with automatic-resync disabled. > Then as continuous writes arrive, move the "recovery_cp" variable forward > towards the end of the array. When it reaches the end we can safely mark the > whole array as ''in-sync'' and forget about diabling auto-resync. > > NeilBrown >That was my thoughts here. I don''t know what state the planned "bitmap of non-sync regions" feature is in, but if and when it is implemented, you would just create the replacement raid1 pair without any synchronisation. Any writes to the pair (such as during a raid5 rebuild) would get written to both disks at the same time. David -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Andrea, On Sat, Nov 23, 2013 at 08:55:08AM +0100, Andrea Mazzoleni wrote:> Hi Piergiorgio, > > > How about par2? How does this work? > I checked the matrix they use, and sometimes it contains some singular > square submatrix. > It seems that in GF(2^16) these cases are just less common. Maybe they > were just unnoticed. > > Anyway, this seems to be an already known problem for PAR2, with an > hypothetical PAR3 fixing it: > > http://sourceforge.net/p/parchive/discussion/96282/thread/d3c6597b/you did a pretty damn good research work! Maybe you should consider to contact them too. I''m not sure if your approach can be extended to GF(2^16), I guess yes, in that case they might be interested too. bye, -- piergiorgio -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/23/2013 1:12 AM, NeilBrown wrote:> On Fri, 22 Nov 2013 21:34:41 -0800 John Williams <jwilliams4200@gmail.com>>> Even a single 8x PCIe 3.0 card has potentially over 7GB/s of bandwidth. >> >> Bottom line is that IO bandwidth is not a problem for a system with >> prudently chosen hardware.Quite right.>> More likely is that you would be CPU limited (rather than bus limited) >> in a high-parity rebuild where more than one drive failed. But even >> that is not likely to be too bad, since Andrea''s single-threaded >> recovery code can recover two drives at nearly 1GB/s on one of my >> machines. I think the code could probably be threaded to achieve a >> multiple of that running on multiple cores. > > Indeed. It seems likely that with modern hardware, the linear write speed > would be the limiting factor for spinning-rust drives.Parity array rebuilds are read-modify-write operations. The main difference from normal operation RMWs is that the write is always to the same disk. As long as the stripe reads and chunk reconstruction outrun the write throughput then the rebuild speed should be as fast as a mirror rebuild. But this doesn''t appear to be what people are experiencing. Parity rebuilds would seem to take much longer. I have always surmised that the culprit is rotational latency, because we''re not able to get a real sector-by-sector streaming read from each drive. If even only one disk in the array has to wait for the platter to come round again, the entire stripe read is slowed down by an additional few milliseconds. For example, in an 8 drive array let''s say each stripe read is slowed 5ms by only one of the 7 drives due to rotational latency, maybe acoustical management, or some other firmware hiccup in the drive. This slows down the entire stripe read because we can''t do parity reconstruction until all chunks are in. An 8x 2TB array with 512KB chunk has 4 million stripes of 4MB each. Reading 4M stripes, that extra 5ms per stripe read costs us (4,000,000 * 0.005)/3600 = 5.56 hours Now consider that arrays typically have a few years on them before the first drive failure. During our rebuild it''s likely that some drives will take a few rotations to return a sector that''s marginal. So this might slow down a stripe read by dozens of milliseconds, maybe a full second. If this happens to multiple drives many times throughout the rebuild it will add even more elapsed time, possibly additional hours. Reading stripes asynchronously or in parallel, which I assume we already do to some degree, can mitigate these latencies to some extent. But I think in the overall picture, things of this nature are what is driving parity rebuilds to dozens of hours for many people. And as I stated previously, when drives reach 10-20TB, this becomes far worse because we''re reading 2-10x as many stripes. And the more drives per array the greater the odds of incurring latency during a stripe read. With a mirror reconstruction we can stream the reads. Though we can''t avoid all of the drive issues above, the total number of hiccups causing latency will be at most 1/7th those of the parity 8 drive array case. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Sat, Nov 23, 2013 at 8:03 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote:> Parity array rebuilds are read-modify-write operations. The main > difference from normal operation RMWs is that the write is always to the > same disk. As long as the stripe reads and chunk reconstruction outrun > the write throughput then the rebuild speed should be as fast as a > mirror rebuild. But this doesn''t appear to be what people are > experiencing. Parity rebuilds would seem to take much longer."This" doesn''t appear to be what SOME people, who have reported issues, are experiencing. Their issues must be examined on a case by case basis. But I, and a number of other people I have talked to or corresponded with, have had mdadm RAID 5 or RAID 6 rebuilds of one drive run at approximately the optimal sequential write speed of the replacement drive. It is not unusual on a reasonably configured system. I don''t know how fast the rebuilds go on the experimental RAID 5 or RAID 6 for btrfs. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Sun, 24 Nov 2013, Stan Hoeppner <stan@hardwarefreak.com> wrote:> I have always surmised that the culprit is rotational latency, because > we''re not able to get a real sector-by-sector streaming read from each > drive. If even only one disk in the array has to wait for the platter > to come round again, the entire stripe read is slowed down by an > additional few milliseconds. For example, in an 8 drive array let''s say > each stripe read is slowed 5ms by only one of the 7 drives due to > rotational latency, maybe acoustical management, or some other firmware > hiccup in the drive. This slows down the entire stripe read because we > can''t do parity reconstruction until all chunks are in. An 8x 2TB array > with 512KB chunk has 4 million stripes of 4MB each. Reading 4M stripes, > that extra 5ms per stripe read costs us > > (4,000,000 * 0.005)/3600 = 5.56 hoursIf that is the problem then the solution would be to just enable read-ahead. Don''t we already have that in both the OS and the disk hardware? The hard- drive read-ahead buffer should at least cover the case where a seek completes but the desired sector isn''t under the heads. RAM size is steadily increasing, it seems that the smallest that you can get nowadays is 1G in a phone and for a server the smallest is probably 4G. On the smallest system that might have an 8 disk array you should be able to use 512M for buffers which allows a read-ahead of 128 chunks.> Now consider that arrays typically have a few years on them before the > first drive failure. During our rebuild it''s likely that some drives > will take a few rotations to return a sector that''s marginal.Are you suggesting that it would be a common case that people just write data to an array and never read it or do an array scrub? I hope that it will become standard practice to have a cron job scrubbing all filesystems.> So this > might slow down a stripe read by dozens of milliseconds, maybe a full > second. If this happens to multiple drives many times throughout the > rebuild it will add even more elapsed time, possibly additional hours.Have you observed such 1 second reads in practice? One thing I''ve considered doing is placing a cheap disk on a speaker cone to test vibration induced performance problems. Then I can use a PC to control the level of vibration in a reasonably repeatable manner. I''d like to see what the limits are for retries. Some years ago a company I worked for had some vibration problems which dropped the contiguous read speed from about 100MB/s to about 40MB/s on some parts of the disk (other parts gave full performance). That was a serious and unusual problem and it only abouty halved the overall speed. -- My Main Blog http://etbe.coker.com.au/ My Documents Blog http://doc.coker.com.au/ -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Piergiorgio, Checking better I found that for PAR3 it was also evaluated a Cauchy matrix, but they preferred to use a RS with FFT in GF(2^16 +1). http://sourceforge.net/mailarchive/forum.php?forum_name=parchive-devel&max_rows=25&style=nested&viewmonth=201006 Note that using a Cauchy matrix for a MDS is something well known. What I did is just to adapt a Cauchy matrix to be compatible with the present Linux kernel implementation. I think it has a value only for Linux or compatible implementations. Ciao, Andrea On Sat, Nov 23, 2013 at 11:10 PM, Piergiorgio Sartor <piergiorgio.sartor@nexgo.de> wrote:> Hi Andrea, > > On Sat, Nov 23, 2013 at 08:55:08AM +0100, Andrea Mazzoleni wrote: >> Hi Piergiorgio, >> >> > How about par2? How does this work? >> I checked the matrix they use, and sometimes it contains some singular >> square submatrix. >> It seems that in GF(2^16) these cases are just less common. Maybe they >> were just unnoticed. >> >> Anyway, this seems to be an already known problem for PAR2, with an >> hypothetical PAR3 fixing it: >> >> http://sourceforge.net/p/parchive/discussion/96282/thread/d3c6597b/ > > you did a pretty damn good research work! > > Maybe you should consider to contact them too. > I''m not sure if your approach can be extended > to GF(2^16), I guess yes, in that case they > might be interested too. > > bye, > > -- > > piergiorgio-- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/23/2013 11:14 PM, John Williams wrote:> On Sat, Nov 23, 2013 at 8:03 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote: > >> Parity array rebuilds are read-modify-write operations. The main >> difference from normal operation RMWs is that the write is always to the >> same disk. As long as the stripe reads and chunk reconstruction outrun >> the write throughput then the rebuild speed should be as fast as a >> mirror rebuild. But this doesn''t appear to be what people are >> experiencing. Parity rebuilds would seem to take much longer. > > "This" doesn''t appear to be what SOME people, who have reported > issues, are experiencing. Their issues must be examined on a case by > case basis.Given what you state below this may very well be the case.> But I, and a number of other people I have talked to or corresponded > with, have had mdadm RAID 5 or RAID 6 rebuilds of one drive run at > approximately the optimal sequential write speed of the replacement > drive. It is not unusual on a reasonably configured system.I freely admit I may have drawn an incorrect conclusion about md parity rebuild performance based on incomplete data. I simply don''t recall anyone stating here in ~3 years that their parity rebuilds were speedy, but quite the opposite. I guess it''s possible that each one of those cases was due to another factor, such as user load, slow CPU, bus bottleneck, wonky disk firmware, backplane issues, etc. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/23/2013 11:19 PM, Russell Coker wrote:> On Sun, 24 Nov 2013, Stan Hoeppner <stan@hardwarefreak.com> wrote: >> I have always surmised that the culprit is rotational latency, because >> we''re not able to get a real sector-by-sector streaming read from each >> drive. If even only one disk in the array has to wait for the platter >> to come round again, the entire stripe read is slowed down by an >> additional few milliseconds. For example, in an 8 drive array let''s say >> each stripe read is slowed 5ms by only one of the 7 drives due to >> rotational latency, maybe acoustical management, or some other firmware >> hiccup in the drive. This slows down the entire stripe read because we >> can''t do parity reconstruction until all chunks are in. An 8x 2TB array >> with 512KB chunk has 4 million stripes of 4MB each. Reading 4M stripes, >> that extra 5ms per stripe read costs us >> >> (4,000,000 * 0.005)/3600 = 5.56 hours > > If that is the problem then the solution would be to just enable read-ahead. > Don''t we already have that in both the OS and the disk hardware? The hard- > drive read-ahead buffer should at least cover the case where a seek completes > but the desired sector isn''t under the heads.I''m not sure if read-ahead would solve such a problem, if indeed this is a possible problem. AFAIK the RAID5/6 drivers process stripes serially, not asynchronously, so I''d think the rebuild may still stall for ms at a time in such a situation.> RAM size is steadily increasing, it seems that the smallest that you can get > nowadays is 1G in a phone and for a server the smallest is probably 4G. > > On the smallest system that might have an 8 disk array you should be able to > use 512M for buffers which allows a read-ahead of 128 chunks. > >> Now consider that arrays typically have a few years on them before the >> first drive failure. During our rebuild it''s likely that some drives >> will take a few rotations to return a sector that''s marginal. > > Are you suggesting that it would be a common case that people just write data > to an array and never read it or do an array scrub? I hope that it will > become standard practice to have a cron job scrubbing all filesystems.Given the frequency of RAID5 double drive failure "save me!" help requests we see on a very regular basis here, it seems pretty clear this is exactly what many users do.>> So this >> might slow down a stripe read by dozens of milliseconds, maybe a full >> second. If this happens to multiple drives many times throughout the >> rebuild it will add even more elapsed time, possibly additional hours. > > Have you observed such 1 second reads in practice?We seem to have regular reports from DIY hardware users intentionally using mismatched consumer drives, as many believe this gives them additional protection against a firmware bug in a given drive model. But then they often see multiple second timeouts causing drives to be kicked, or performance to be slow, because of the mismatched drives. In my time on this list, it seems pretty clear that the vast majority of posters use DIY hardware, not matched, packaged, tested solutions from the likes of Dell, HP, IBM, etc. Some of the things I''ve speculated about in my last few posts could very well occur, and indeed be caused by, ad hoc component selection and system assembly. Obviously not in all DIY cases, but probably many. -- Stan> One thing I''ve considered doing is placing a cheap disk on a speaker cone to > test vibration induced performance problems. Then I can use a PC to control > the level of vibration in a reasonably repeatable manner. I''d like to see > what the limits are for retries. > > Some years ago a company I worked for had some vibration problems which > dropped the contiguous read speed from about 100MB/s to about 40MB/s on some > parts of the disk (other parts gave full performance). That was a serious and > unusual problem and it only abouty halved the overall speed.-- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Sun, Nov 24, 2013 at 1:44 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote: <SNIP>>> Are you suggesting that it would be a common case that people just write data >> to an array and never read it or do an array scrub? I hope that it will >> become standard practice to have a cron job scrubbing all filesystems. > > Given the frequency of RAID5 double drive failure "save me!" help > requests we see on a very regular basis here, it seems pretty clear this > is exactly what many users do.Thank you for reminding me to set this up. Sunday afternoons at 2:30 is now my scrub time. Cheers, Mark -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 24-11-13 22:13, Stan Hoeppner wrote:> I freely admit I may have drawn an incorrect conclusion about md > parity rebuild performance based on incomplete data. I simply don''t > recall anyone stating here in ~3 years that their parity rebuilds were > speedy, but quite the opposite. I guess it''s possible that each one of > those cases was due to another factor, such as user load, slow CPU, > bus bottleneck, wonky disk firmware, backplane issues, etc.The other part to this is, that the list sees what goes wrong. The cases where a replacement goes smoothly do not reach the list Cheers, Rudy -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Stan Hoeppner wrote:> On 11/23/2013 11:14 PM, John Williams wrote: >> On Sat, Nov 23, 2013 at 8:03 PM, Stan Hoeppner <stan@hardwarefreak.com> >> wrote:<snip>> >> But I, and a number of other people I have talked to or corresponded >> with, have had mdadm RAID 5 or RAID 6 rebuilds of one drive run at >> approximately the optimal sequential write speed of the replacement >> drive. It is not unusual on a reasonably configured system. > > I freely admit I may have drawn an incorrect conclusion about md parity > rebuild performance based on incomplete data. I simply don''t recall > anyone stating here in ~3 years that their parity rebuilds were speedy, > but quite the opposite. I guess it''s possible that each one of those > cases was due to another factor, such as user load, slow CPU, bus > bottleneck, wonky disk firmware, backplane issues, etc. >Well, there''s also the issue of selection bias - people come to the list and complain when their RAID is taking forever to resync. People generally don''t come to the list and complain when their RAID resyncs quickly and without issues. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/24/2013 5:53 PM, Alex Elsayed wrote:> Stan Hoeppner wrote: > >> On 11/23/2013 11:14 PM, John Williams wrote: >>> On Sat, Nov 23, 2013 at 8:03 PM, Stan Hoeppner <stan@hardwarefreak.com> >>> wrote: > <snip> >> >>> But I, and a number of other people I have talked to or corresponded >>> with, have had mdadm RAID 5 or RAID 6 rebuilds of one drive run at >>> approximately the optimal sequential write speed of the replacement >>> drive. It is not unusual on a reasonably configured system. >> >> I freely admit I may have drawn an incorrect conclusion about md parity >> rebuild performance based on incomplete data. I simply don''t recall >> anyone stating here in ~3 years that their parity rebuilds were speedy, >> but quite the opposite. I guess it''s possible that each one of those >> cases was due to another factor, such as user load, slow CPU, bus >> bottleneck, wonky disk firmware, backplane issues, etc. >> > > Well, there''s also the issue of selection bias - people come to the list and"Selection bias" would infer I''m doing some kind of formal analysis, which is obviously not the case, though I do understand the point you''re making.> complain when their RAID is taking forever to resync. People generally don''t > come to the list and complain when their RAID resyncs quickly and without > issues.When folks report problems on linux-raid it is commonplace for others to reply that the same feature works fine for them, that the problem may be configuration specific, etc. When people have reported slow RAID5/6 rebuilds in the past, and these were not always reported in direct help requests but as "me too" posts, I don''t recall others saying their parity rebuilds are speedy. I''m not saying nobody ever has, simply that I don''t recall such. Which is why I''ve been under the impression that parity rebuilds are generally slow for everyone. I wish I had hardware available to perform relevant testing. It would be nice to have some real data on this showing apples to apples rebuild times for the various RAID levels on the same hardware. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Mon, 25 Nov 2013, Stan Hoeppner <stan@hardwarefreak.com> wrote:> > If that is the problem then the solution would be to just enable > > read-ahead. Don''t we already have that in both the OS and the disk > > hardware? The hard- drive read-ahead buffer should at least cover the > > case where a seek completes but the desired sector isn''t under the > > heads. > > I''m not sure if read-ahead would solve such a problem, if indeed this is > a possible problem. AFAIK the RAID5/6 drivers process stripes serially, > not asynchronously, so I''d think the rebuild may still stall for ms at a > time in such a situation.For a RAID block device (such as Linux software RAID) read-ahead should work well. For a RAID type configuration managed by the filesystem where you might have different RAID levels in the same filesystem it might not be possible. It would be a nice feature to have RAID-0 for unimportant files and RAID-1 or RAID-6 for important files on the same filesystem. But that type of thing would really complicate RAID rebuild.> >> So this > >> might slow down a stripe read by dozens of milliseconds, maybe a full > >> second. If this happens to multiple drives many times throughout the > >> rebuild it will add even more elapsed time, possibly additional hours. > > > > Have you observed such 1 second reads in practice? > > We seem to have regular reports from DIY hardware users intentionally > using mismatched consumer drives, as many believe this gives them > additional protection against a firmware bug in a given drive model. > But then they often see multiple second timeouts causing drives to be > kicked, or performance to be slow, because of the mismatched drives.I''ve had that happen in my own RAID-1 and had reports of the same from other people. My problem was that one of the disks was a WD "Green" drive that aggressively parked the heads. When I replaced the array the disk in question had parked it''s heads about a million times more than it was supposed to but it still worked. I think that if you want to buy two different disks from your local discount store then you are likely to end up with one that''s not well suited to server use because they often don''t have more than two types of disks on offer. -- My Main Blog http://etbe.coker.com.au/ My Documents Blog http://doc.coker.com.au/ -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Stan Hoeppner wrote: <snip<> "Selection bias" would infer I''m doing some kind of formal analysis, > which is obviously not the case, though I do understand the point you''re > making.Nope - the only thing ''selection bias'' says is that inferences drawn from a non-representative dataset will also be non-representative. Drawing an inference of ''raid rebuilds are slow'' from a mailing list people come to when they need to report an issue will overemphasize problems and deemphasize cases where everything is working properly. Also, no matter how quickly a rebuild goes, it''s still longer than people would like. No matter how quickly it finishes, you''re still in a stressed state where something has _already_ gone wrong and your parachute is now in use and unable to save you a second time. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 24/11/13 22:13, Stan Hoeppner wrote:> On 11/23/2013 11:14 PM, John Williams wrote: >> On Sat, Nov 23, 2013 at 8:03 PM, Stan Hoeppner <stan@hardwarefreak.com> wrote: >> >>> Parity array rebuilds are read-modify-write operations. The main >>> difference from normal operation RMWs is that the write is always to the >>> same disk. As long as the stripe reads and chunk reconstruction outrun >>> the write throughput then the rebuild speed should be as fast as a >>> mirror rebuild. But this doesn''t appear to be what people are >>> experiencing. Parity rebuilds would seem to take much longer. >> >> "This" doesn''t appear to be what SOME people, who have reported >> issues, are experiencing. Their issues must be examined on a case by >> case basis. > > Given what you state below this may very well be the case. > >> But I, and a number of other people I have talked to or corresponded >> with, have had mdadm RAID 5 or RAID 6 rebuilds of one drive run at >> approximately the optimal sequential write speed of the replacement >> drive. It is not unusual on a reasonably configured system. > > I freely admit I may have drawn an incorrect conclusion about md parity > rebuild performance based on incomplete data. I simply don''t recall > anyone stating here in ~3 years that their parity rebuilds were speedy, > but quite the opposite. I guess it''s possible that each one of those > cases was due to another factor, such as user load, slow CPU, bus > bottleneck, wonky disk firmware, backplane issues, etc. >Maybe this is just reporting bias - people are quick to post about problems such as slow rebuilds, but very seldom send a message saying everything worked perfectly! There /are/ reasons why parity raid rebuilds are going to be slower than mirror rebuilds - delays on one disk reading is one issue, and I expect that simultaneous use of the array for normal work will have more impact on parity raid rebuild times than on a mirror array (certainly compared to raid10 with multiple pairs). I just don''t think it is quite as bad as you think. -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 25/11/13 03:14, Russell Coker wrote:> On Mon, 25 Nov 2013, Stan Hoeppner <stan@hardwarefreak.com> wrote: >>> If that is the problem then the solution would be to just enable >>> read-ahead. Don''t we already have that in both the OS and the disk >>> hardware? The hard- drive read-ahead buffer should at least cover the >>> case where a seek completes but the desired sector isn''t under the >>> heads. >> >> I''m not sure if read-ahead would solve such a problem, if indeed this is >> a possible problem. AFAIK the RAID5/6 drivers process stripes serially, >> not asynchronously, so I''d think the rebuild may still stall for ms at a >> time in such a situation. > > For a RAID block device (such as Linux software RAID) read-ahead should work > well. For a RAID type configuration managed by the filesystem where you might > have different RAID levels in the same filesystem it might not be possible. > > It would be a nice feature to have RAID-0 for unimportant files and RAID-1 or > RAID-6 for important files on the same filesystem. But that type of thing > would really complicate RAID rebuild. >I think btrfs is planning to have such features - different files can have different raid types. It certainly supports different raid levels for metadata and file data. But it is definitely a feature you want on the filesystem level, rather than the raid block device level. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, Nov 22, 2013 at 07:12:39PM +1100, Russell Coker wrote:> > On Thu, 21 Nov 2013 18:30:49 Stan Hoeppner wrote: > > I suggest that anyone in the future needing fast random write IOPS is > > going to move those workloads to SSD, which is steadily increasing in > > capacity. And I suggest anyone building arrays with 10-20TB drives > > isn''t in need of fast random write IOPS. > > Traditionally SCSI/SAS disks have tended to be a lot smaller than IDE/SATA > disks. Now Dell has just started offering 2TB SAS disks while the largest SATA > disks that they sell (In Australia on PowerEdge T110 servers at least) are > also 2TB. Presumably RAID recovery time was one factor that made > manufacturers not bother with making larger SCSI/SAS disks in the past. >Dell has been selling 4TB SAS disks for a long time already.. one year, or something like that. -- Pasi -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 21/11/2013 21:31, David Brown wrote:> On 21/11/13 21:05, Piergiorgio Sartor wrote: >> Having a multi parity RAID allows to check >> even which disk. >> This would provide the user with a more >> comprehensive (I forgot the spelling) >> information. >> >> Of course, since we are there, we can >> also give the option to fix it. >> This would be much likely a "fsck". > If this can all be done to give the user an informed choice, then it > sounds good. > > One issue here is whether the check should be done with the filesystem > mounted and in use, or only off-line. If it is off-line then it will > mean a long down-time while the array is checked - but if it is online, > then there is the risk of confusing the filesystem and caches by > changing the data. >Non-existent issue imho, because if that stripe is changing, any error will be corrected (overwritten), at least on the data disks (parity can still be wrong if a shortcut-rmw method is used). So you perform fsck for filesystem and data comes out good because any error has been overwritten already, and fsck also returns noerror. Not useful. You have to consider only the case where the array check is performed online, and the stripe does not change in the meanwhile, that means it does not change for a long time, enough for you to complete all the checks. debugfs techniques can tell you what filesystem element corresponds to a certain block number, and this can be done online, with the filesystem mounted. I don''t understand the thing you say about the caches. Caches are not an obstacle for current "check" operation, so they also won''t be a problem for the new improved check operation which you are discussing. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
This is great stuff. Now, how can we get this into btrfs and md? On Wed, Nov 20, 2013 at 1:23 PM, Andrea Mazzoleni <amadvance@gmail.com> wrote:> Hi, > >> First, create a 3 by 6 cauchy matrix, using x_i = 2^-i, and y_i = 0 for i=0, and y_i = 2^i for other i. >> In this case: x = { 1, 142, 71, 173, 216, 108 } y = { 0, 2, 4). The cauchy matrix is: >> >> 1 2 4 8 16 32 >> 244 83 78 183 118 47 >> 167 39 213 59 153 82 >> >> Divide row 2 by 244 and row 3 by 167. Then extend it with a row of ones on top and it''s still MDS, >> and that''s the code for m=4, with RAID-6 as a subset. Very nice! > > You got it Jim! > > Thanks, > Andrea > -- > To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html-- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Late reply. This one got lost in the flurry of activity... On 11/22/2013 7:24 AM, David Brown wrote:> On 22/11/13 09:38, Stan Hoeppner wrote: >> On 11/21/2013 3:07 AM, David Brown wrote: >> >>> For example, with 20 disks at 1 TB each, you can have: >>...>> Maximum: >> >> RAID 10 = 10 disk redundancy >> RAID 15 = 11 disk redundancy > > 12 disks maximum (you have 8 with data, the rest are mirrors, parity, or > mirrors of parity). > >> RAID 16 = 12 disk redundancy > > 14 disks maximum (you have 6 with data, the rest are mirrors, parity, or > mirrors of parity).We must follow different definitions of "redundancy". I view redundancy as the number of drives that can fail without taking down the array. In the case of the above 20 drive RAID15 that maximum is clearly 11 drives-- one of every mirror and both of one mirror can fail. The 12th drive failure kills the array.>> Range: >> >> RAID 10 = 1-10 disk redundancy >> RAID 15 = 3-11 disk redundancy >> RAID 16 = 5-12 disk redundancy > > Yes, I know these are the minimum redundancies. But that''s a vital > figure for reliability (even if the range is important for statistical > averages). When one disk in a raid10 array fails, your main concern is > about failures or URE''s in the other half of the pair - it doesn''t help > to know that another nine disks can "safely" fail too.Knowing this is often critical from an architectural standpoint David. It is quite common to create the mirrors of a RAID10 across two HBAs and two JBOD chassis. Some call this "duplexing". With RAID10 you know you can lose one HBA, one cable, one JBOD (PSU, expander, etc) and not skip a beat. "RAID15" would work the same in this scenario. This architecture is impossible with RAID5/6. Any of the mentioned failures will kill the array. -- Stan -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, 28 Nov 2013, Stan Hoeppner <stan@hardwarefreak.com> wrote:> We must follow different definitions of "redundancy". I view redundancy > as the number of drives that can fail without taking down the array. In > the case of the above 20 drive RAID15 that maximum is clearly 11 > drives-- one of every mirror and both of one mirror can fail. The 12th > drive failure kills the array.It seems to me that the more useful number is how many disks can randomly die while there is a guarantee that no data is lost. While with a 20 disk RAID-15 you could get really lucky and have it still work after 11 disks have been lost you could lose it all after 4 disks have been lost. Given that drive failures won''t be entirely random (when one disk dies it puts more load on it''s mirror) having a 20 disk RAID-15 entirely fail after 4 or 5 disk failures is probably more likely than having it keep working after 11 failures. But as far as I can determine in the 20 disk RAID-15 array in question the "redundancy" would be 11 disks because once you have lost that many disks the usable capacity would be equal to the number of running disks multiplied by their capacity. For simpler RAID configurations the "redundancy" equals the maximum number of disks that can randomly fail without losing any data while with nested RAIDs that''s not the case. -- My Main Blog http://etbe.coker.com.au/ My Documents Blog http://doc.coker.com.au/ -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 28/11/13 08:16, Stan Hoeppner wrote:> Late reply. This one got lost in the flurry of activity... > > On 11/22/2013 7:24 AM, David Brown wrote: >> On 22/11/13 09:38, Stan Hoeppner wrote: >>> On 11/21/2013 3:07 AM, David Brown wrote: >>> >>>> For example, with 20 disks at 1 TB each, you can have: >>> > ... >>> Maximum: >>> >>> RAID 10 = 10 disk redundancy >>> RAID 15 = 11 disk redundancy >> >> 12 disks maximum (you have 8 with data, the rest are mirrors, parity, or >> mirrors of parity). >> >>> RAID 16 = 12 disk redundancy >> >> 14 disks maximum (you have 6 with data, the rest are mirrors, parity, or >> mirrors of parity). > > We must follow different definitions of "redundancy". I view redundancy > as the number of drives that can fail without taking down the array. In > the case of the above 20 drive RAID15 that maximum is clearly 11 > drives-- one of every mirror and both of one mirror can fail. The 12th > drive failure kills the array. >No, we have the same definitions of redundancy - just different definitions of basic arithmetic. Your definition is a bit more common! My error was actually in an earlier email, when I listed the usable capacities of different layouts for 20 x 1TB drive. I wrote:> raid10 = 10TB, 1 disk redundancy > raid15 = 8TB, 3 disk redundancy > raid16 = 6TB, 5 disk redundancyOf course, it should be: raid10 = 10TB, 1 disk redundancy raid15 = 9TB, 3 disk redundancy raid16 = 8TB, 5 disk redundancy So it is your fault for not spotting my earlier mistake :-)>>> Range: >>> >>> RAID 10 = 1-10 disk redundancy >>> RAID 15 = 3-11 disk redundancy >>> RAID 16 = 5-12 disk redundancy >> >> Yes, I know these are the minimum redundancies. But that''s a vital >> figure for reliability (even if the range is important for statistical >> averages). When one disk in a raid10 array fails, your main concern is >> about failures or URE''s in the other half of the pair - it doesn''t help >> to know that another nine disks can "safely" fail too. > > Knowing this is often critical from an architectural standpoint David. > It is quite common to create the mirrors of a RAID10 across two HBAs and > two JBOD chassis. Some call this "duplexing". With RAID10 you know you > can lose one HBA, one cable, one JBOD (PSU, expander, etc) and not skip > a beat. "RAID15" would work the same in this scenario. >That is absolutely true, and I agree that it is very important when setting up big arrays. You have to make decisions like where you split your raid1 pairs - putting them on different controllers/chassis means you can survive the loss of a whole half of the system. On the other hand, putting them on the same controller could mean hardware raid1 is more efficient and you don''t need to duplicate the traffic over the higher level interfaces. But here we are looking at one specific class of failures - hard disk failures (including complete disk failure and URE''s). For that, the redundancy is the number of disks that can fail without data loss, assuming the worst possible combination of failures. And given the extra stress on the disks during degraded access or rebuilds, "bad" combinations are more likely than "good" combinations. So I think it is of little help to say that a 20 disk raid 15 can survive up to 11 disk failures. It is far more interesting to say that it can survive any 3 random disk failures, and (if connected as you describe with two controllers and chassis) it can also survive the complete failure of a chassis or controller while still retaining a one disk redundancy. As a side issue here, I wonder if a write intent bitmap can be used for a chassis failure so that when the chassis is fixed (the controller card replaced, the cable re-connected, etc.) the disks inside can be brought up to sync again without a full rebuild.> This architecture is impossible with RAID5/6. Any of the mentioned > failures will kill the array. >Yes. -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Stan Hoeppner wrote:> Late reply. This one got lost in the flurry of activity... ><snip>> > We must follow different definitions of "redundancy". I view redundancy > as the number of drives that can fail without taking down the array. In > the case of the above 20 drive RAID15 that maximum is clearly 11 > drives-- one of every mirror and both of one mirror can fail. The 12th > drive failure kills the array.IMHO, ''redundancy'' is not the important thing here, and may conflate two things - ''how much storage is spent on things other than my data (mirrors, parity)'' [storage efficiency] and ''if the universe is trying to screw me over, how many disk losses can I survive?'' [failure resilience] Your 11 disks number is best-case, but quicksort has taught me to always look at average-case and worst-case as well. What you describe has very good best-case failure resilience, but its worst-case resilience is poorer. It has better best-case, average-case, *and* worst-case performance, but has considerably worse storage efficiency. All of those need to be weighed in deciding which to use; raid 15 being ''just better'' isn''t something that can be claimed. It depends on the workload. <snip>> Knowing this is often critical from an architectural standpoint David. > It is quite common to create the mirrors of a RAID10 across two HBAs and > two JBOD chassis. Some call this "duplexing". With RAID10 you know you > can lose one HBA, one cable, one JBOD (PSU, expander, etc) and not skip > a beat. "RAID15" would work the same in this scenario. > > This architecture is impossible with RAID5/6. Any of the mentioned > failures will kill the array.And again, these address different issues. For one, there''s multipath - with dual-ported drives, multipath, etc. you can tolerate HBA failure; that renders part of the issue something of a canard. Adding a second JBOD is really not an apples-to-apples comparison, either - it''s not likely to be a useful configuration in the same situations as lend themselves to parity >> 2. Certainly, no-one is advocating that support for RAID 10 be removed, and mdadm is certainly capable of handling a manually- created RAID 15 today. -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html