Hi, sorry if someone else already asked this, im just trying to understand, answers for this kind of subject are hard to find. First of all my real problem is quite easy, im looking for a way to setup a file server, but in witch i would have these features : - having at least 10 physical drives - having not any data redundancy over all physical drive (spanned,raid linear mode) - having all physical disk using the same "volume" (logical path) with the full available disks space - having ability to remove/add physical disk in this "volume" - having ability to access this volume regardless of offline physical drives, and having access to the data that are remaining on the still online physical drives (of course for data from offline drives, cant be access... but would re-access them when plugged back on and re-added to "volume"...) - i dont care of hotplug features Does ZFS will do that ?? If it is possible, then the question would be, how the data would be written to disks? the only match i founded are the windows xp spanned mode or raid linear mode, for both, the files are written to the first available drive in the array, and data are filled on the drive until drive full and proced to next drive until space is totaly full in the array. For raid linear, as a disk is missing the whole array die, and from what i understood no way of getting remaining disks data, for windows, the file system is on the first disk, and you can rebuild some of data if the first disk not off (rebuild, but cant realy acces them, the volume remain offline, all you can do is use software for backup data), the main problem is that windows suck and dealing with such setup is a BIG pain and BIG loss of times. If someone could help me understanding how to make such a system with lot of disk, no data redundancy, simple management and data preservation along the safe drives, by ZFS or any other software method it would definitly make my day. Thx in advance for reply. ToTo This message posted from opensolaris.org
So... you want no data redundancy, but want to be able to access your data even though you''ve lost an arbitrary number of drives? No, ZFS will not do this - it''s not PsychicFS (as Bill would say). We have some enhancements coming soon which will implicitly mirror metadata, but not user data. This should ensure that your data is at least discoverable if you lose a disk in a dynamic stripe, but you won''t be able to actually access the underlying file contents. Moreover, which pieces are inaccessible are completely unpredictable, since the blocks could have been allocated anywhere. Lose the right number of disks and you may lose the root of your filesystem (or worse yet, storage pool). Out of curiosity, why is this setup beneficial? Why not use RAID-Z or some other form of redundancy? Is the cost of one disk out of N too large? More importantly, is the cost of said disk worth more than all of your data? - Eric -- Eric Schrock, Solaris Kernel Development http://blogs.sun.com/eschrock
>So... you want no data redundancy, but want to be able to access your >data even though you''ve lost an arbitrary number of drives?i think one drive fail at a time is sufficent, not several (drives are not dieing alltogether statisticaly in 99% of case of disks crashs)>Out of curiosity, why is this setup beneficial?Having a one single only logical drive like for p2p data holding (careless of data loss/integrity)>Is the cost of one disk out of N too >large?in this case yes, since raid5(or Z whatever redundancy) would have some perf issues if i dont do hardware setup correctly, so not only one disk price (linear mode dont access through all disks at the same time, one disk by one, no perfs lacks as a raid would do)>More importantly, is the cost of said disk > worth more than all > of your data?in that case definitly yes (if when u say all data, u would mean all data lost from the crashed drive) thx for having reply, ive founded a way to get trought that with EVMS, the only lack is when disk0 in array die, all die, BUT i can have a disk fail, replace the drive (like a empty one) in the array, then i can fsck who will fix the lost files and having space available, previous disks data are still there. This message posted from opensolaris.org
On Sun, Jan 29, 2006 at 09:06:16AM -0800, Titi Caca wrote:> > in this case yes, since raid5(or Z whatever redundancy) would have > some perf issues if i dont do hardware setup correctly, so not only > one disk price (linear mode dont access through all disks at the same > time, one disk by one, no perfs lacks as a raid would do)Have you actually benchmarked this? RAID-Z is significantly faster than any other software RAID solution out there, since we can avoid the read-modify-write problem of avoiding the write hole.> thx for having reply, ive founded a way to get trought that with EVMS, > the only lack is when disk0 in array die, all die, BUT i can have a > disk fail, replace the drive (like a empty one) in the array, then i > can fsck who will fix the lost files and having space available, > previous disks data are still there.With the metadata replication I mentioned earlier, you would get this with ZFS, only you wouldn''t have to do any fsck or "fixing". But this feature isn''t in the current bits - it should be available within the next 6 weeks. - Eric -- Eric Schrock, Solaris Kernel Development http://blogs.sun.com/eschrock
>Have you actually benchmarked this? RAID-Z is significantly faster than >any other software RAID solution out there, since we can avoid the >read-modify-write problem of avoiding the write hole.sry ive only monitored raid5 perfs witch are bad in my case>With the metadata replication I mentioned earlier, you would get this >with ZFS, only you wouldn''t have to do any fsck or "fixing". But this >feature isn''t in the current bits - it should be available within the >next 6 weeks.will try it soon so :) thx So if i dont do a backup of "files lists" before the disk crash, the files would disapear never telling me who was killed in the process. Last thing, in metadata replication with ZFS, does the disk0 of the array is mandatory for the "volume" to work like with EVMS? or all disk are containing the full info for making every disk working? This message posted from opensolaris.org
On Mon, Jan 30, 2006 at 02:44:23AM -0800, Titi Caca wrote:> > will try it soon so :) thx So if i dont do a backup of "files lists" > before the disk crash, the files would disapear never telling me who > was killed in the process.No, an upcoming feature of ZFS is a ''persistent error log'' which tracks all unrecoverable errors within a pool. As you try to access the data, they would be logged here. Alternatively, you could run ''zpool scrub'', which would touch all your data and log any failures as well. Note that this is an online operation, unlike a downtime-inducing fsck(1M). Then, you simply run ''zpool status -v'' to get a list of all errors, with dataset and file names.> Last thing, in metadata replication with ZFS, does the disk0 of the > array is mandatory for the "volume" to work like with EVMS? or all > disk are containing the full info for making every disk working?The data is spread across all disks. The only issue might be where the uberblock was written. In the past, it has been assumed that losing a single disk makes an unreplicated pool unusable. We used to write the uberblock round robin, but I forget if we changed it to always write to the first vdev. With mirrored metadata, this is no longer true. Presumably, this change will also involve implicitly mirroring the uberblock among several disks. Assuming that is the case, there is no "special" disk which you must have in order to access the pool. - Eric -- Eric Schrock, Solaris Kernel Development http://blogs.sun.com/eschrock