Interesting flash technology overview and SSD review here:
http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3403
and another review here:
http://www.tomshardware.com/reviews/Intel-x25-m-SSD,2012.html
Regards,
--
Al Hopper Logical Approach Inc,Plano,TX al at logical-approach.com
Voice: 972.379.2133 Timezone: US CDT
OpenSolaris Governing Board (OGB) Member - Apr 2005 to Mar 2007
http://www.opensolaris.org/os/community/ogb/ogb_2005-2007/
On Wed, 10 Sep 2008, Al Hopper wrote:> Interesting flash technology overview and SSD review here: > > http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3403 > and another review here: > http://www.tomshardware.com/reviews/Intel-x25-m-SSD,2012.htmlThese seem like regurgitations of the same marketing drivel that you notified us about before. These Intel products are assembled in China based on non-Intel FLASH components (from Micron). There is little reason to believe that Intel will "corner the market" due to having an aggressive marketing department. There are other companies in the business who may seem oddly silent compared with Intel/Micron, but enjoy a vastly larger share of the FLASH market. These reviews continue their apples/oranges comparison by comparing cheap lowest-grade desktop/laptop drives with the expensive Intel SSD drives. The hard drive performance specified is for low-grade consumer drives rather than enterprise drives. The hard drive reliability specified is for low-grade consumer drives rather than enterprise drives. The table at Tom''s Hardware talks about 160GB SSD drives which are not even announced. The SLC storage sizes are still quite tiny. The wear leveling algorithm ensures that the drive starts losing its memory in all locations at about the same time. RAID does not really help much here for reliability since RAID systems are usually comprised of the same devices installed at the same time and seeing identical write activity. RAID works due to failures being random. If the failures are not random (i.e. all drives start reporting read errors at once) then RAID does not really help. Hopefully the integration with the OS is sufficient that the user knows it is time to change out the drive before it is too late to salvage the data. Write performance to SSDs is not all it is cracked up to be. Buried in the AnandTech writeup, there is mention that while 4K can be written at once, 512KB needs to be erased at once. This means that write performance to an empty device will seem initially pretty good, but then it will start to suffer as 512KB regions need to be erased to make space for more writes. ZFS''s COW scheme will intially be fast, but then the writes will slow after all blocks on the device have been written to before. Since writing to a used drive incurs additional latency, the device will need to buffer writes in RAM so that it returns to the user faster. This may increase the chance of data loss due to power failure. Bob =====================================Bob Friesenhahn bfriesen at simple.dallas.tx.us, http://www.simplesystems.org/users/bfriesen/ GraphicsMagick Maintainer, http://www.GraphicsMagick.org/
On Sep 10, 2008, at 11:40 AM, Bob Friesenhahn wrote:> > Write performance to SSDs is not all it is cracked up to be. Buried > in the AnandTech writeup, there is mention that while 4K can be > written at once, 512KB needs to be erased at once. This means that > write performance to an empty device will seem initially pretty good, > but then it will start to suffer as 512KB regions need to be erased to > make space for more writes.That assumes that one doesn''t code up the system to batch up erases prior to writes. ...> returns to the user faster. This may increase the chance of data loss > due to power failure.Presumably anyone deft enough to design such an enterprise grade device will be able to provide enough super-capacitor (or equivalent) to ensure that DRAM is flushed to SSD before anything bad happens. Clever use of such devices in L2ARC and slog ZFS configurations (or moral equivalents in other environments) is pretty much the only affordable way (vs. huge numbers of spindles) to bridge the gap between rotating rust and massively parallel CPUs. One imagines that Intel will go back to fabbing their own at some point; that is closer to their usual business model than OEMing other people''s parts ;> -- Keith H. Bierman khbkhb at gmail.com | AIM kbiermank 5430 Nassau Circle East | Cherry Hills Village, CO 80113 | 303-997-2749 <speaking for myself*> Copyright 2008
Bob Friesenhahn wrote:> On Wed, 10 Sep 2008, Al Hopper wrote: > > >> Interesting flash technology overview and SSD review here: >> >> http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3403 >> and another review here: >> http://www.tomshardware.com/reviews/Intel-x25-m-SSD,2012.html >> > > These seem like regurgitations of the same marketing drivel that you > notified us about before. > > These Intel products are assembled in China based on non-Intel FLASH > components (from Micron). There is little reason to believe that > Intel will "corner the market" due to having an aggressive marketing > department. There are other companies in the business who may seem > oddly silent compared with Intel/Micron, but enjoy a vastly larger > share of the FLASH market. >Intel and Micron have a joint venture for doing the flash SSDs. For some reason, Intel''s usually excellent marketing team wasn''t involved in naming the JV, so it is called "IM Flash Technologies"... boring :-) http://www.imftech.com/ Samsung is another major vendor, rumored to be trying to buy Sandisk, but it ain''t over ''til its over... might be a JV opportunity, too. http://www.ft.com/cms/s/2/eb1f748e-7f34-11dd-a3da-000077b07658.html> These reviews continue their apples/oranges comparison by comparing > cheap lowest-grade desktop/laptop drives with the expensive Intel SSD > drives. The hard drive performance specified is for low-grade > consumer drives rather than enterprise drives. The hard drive > reliability specified is for low-grade consumer drives rather than > enterprise drives. The table at Tom''s Hardware talks about 160GB SSD > drives which are not even announced. > > The SLC storage sizes are still quite tiny. The wear leveling > algorithm ensures that the drive starts losing its memory in all > locations at about the same time. RAID does not really help much here > for reliability since RAID systems are usually comprised of the same > devices installed at the same time and seeing identical write > activity. RAID works due to failures being random. If the failures > are not random (i.e. all drives start reporting read errors at once) > then RAID does not really help. Hopefully the integration with the OS > is sufficient that the user knows it is time to change out the drive > before it is too late to salvage the data. >I think the market segments are becoming more solidified. There is clearly a low-cost consumer market. But there is also a large, unsatisfied demand for enterprise-class SSDs. Intel has already announced an SLC based "extreme" product line. Brian''s blog seems to be one of the best distilled descriptions I''ve seen: http://www.edn.com/blog/400000040/post/360032036.html -- richard> Write performance to SSDs is not all it is cracked up to be. Buried > in the AnandTech writeup, there is mention that while 4K can be > written at once, 512KB needs to be erased at once. This means that > write performance to an empty device will seem initially pretty good, > but then it will start to suffer as 512KB regions need to be erased to > make space for more writes. ZFS''s COW scheme will intially be fast, > but then the writes will slow after all blocks on the device have been > written to before. Since writing to a used drive incurs additional > latency, the device will need to buffer writes in RAM so that it > returns to the user faster. This may increase the chance of data loss > due to power failure. > > Bob > =====================================> Bob Friesenhahn > bfriesen at simple.dallas.tx.us, http://www.simplesystems.org/users/bfriesen/ > GraphicsMagick Maintainer, http://www.GraphicsMagick.org/ > > _______________________________________________ > zfs-discuss mailing list > zfs-discuss at opensolaris.org > http://mail.opensolaris.org/mailman/listinfo/zfs-discuss >
On Wed, 10 Sep 2008, Keith Bierman wrote:>> written at once, 512KB needs to be erased at once. This means that >> write performance to an empty device will seem initially pretty good, >> but then it will start to suffer as 512KB regions need to be erased to >> make space for more writes. > > That assumes that one doesn''t code up the system to batch up erases prior to > writes.Is the notion of block "erase" even exposed via SATA/SCSI protocols? Maybe it is for CD/DVD type devices. This is something that only the device itself would be aware of. Only the device knows if the block has been used before. Only the device knows the block of physical storage which will be used for the write. The device does not know what can be erased before it sees a (over) write request and if the write request is for a smaller size, then existing data needs to be moved (for leveling) or buffered and written back to the same locations. This means that 512KB needs to be erased and re-written.> Presumably anyone deft enough to design such an enterprise grade device will > be able to provide enough super-capacitor (or equivalent) to ensure that DRAM > is flushed to SSD before anything bad happens.That is reasonable. It adds to product cost and size though. Super-capacitors are not super-small. Bob =====================================Bob Friesenhahn bfriesen at simple.dallas.tx.us, http://www.simplesystems.org/users/bfriesen/ GraphicsMagick Maintainer, http://www.GraphicsMagick.org/
On Sep 10, 2008, at 12:37 PM, Bob Friesenhahn wrote:> On Wed, 10 Sep 2008, Keith Bierman wrote: > >>> written at once, 512KB needs to be erased at once. This means that >>> write performance to an empty device will seem initially pretty >>> good, >>> but then it will start to suffer as 512KB regions need to be >>> erased to >>> make space for more writes. >> >> That assumes that one doesn''t code up the system to batch up >> erases prior to writes. > > Is the notion of block "erase" even exposed via SATA/SCSI > protocols? Maybe it is for CD/DVD type devices. > > This is something that only the device itself would be aware of. > Only the device knows if the block has been used before.A conspiracy between the device and a savvy host is sure to emerge ;>> ... > That is reasonable. It adds to product cost and size though. Super- > capacitors are not super-small. >True, but for enterprise class devices they are sufficiently small. Laptops will have a largish battery and won''t need the caps ;> Desktops will be on their own. -- Keith H. Bierman khbkhb at gmail.com | AIM kbiermank 5430 Nassau Circle East | Cherry Hills Village, CO 80113 | 303-997-2749 <speaking for myself*> Copyright 2008
On Wed, 10 Sep 2008, Keith Bierman wrote:>> ... >> That is reasonable. It adds to product cost and size though. >> Super-capacitors are not super-small. >> > True, but for enterprise class devices they are sufficiently small. Laptops > will have a largish battery and won''t need the caps ;> Desktops will be on > their own.The Intel SSDs are still not advertised as enterprise class devices. Bob =====================================Bob Friesenhahn bfriesen at simple.dallas.tx.us, http://www.simplesystems.org/users/bfriesen/ GraphicsMagick Maintainer, http://www.GraphicsMagick.org/
On Wed, Sep 10, 2008 at 1:46 PM, Bob Friesenhahn <bfriesen at simple.dallas.tx.us> wrote:> On Wed, 10 Sep 2008, Keith Bierman wrote: >>> ... >>> That is reasonable. It adds to product cost and size though. >>> Super-capacitors are not super-small. >>> >> True, but for enterprise class devices they are sufficiently small. Laptops >> will have a largish battery and won''t need the caps ;> Desktops will be on >> their own. > > The Intel SSDs are still not advertised as enterprise class devices.This is clear from what Intel have already said and what the reviewers have "regurgitated". SSDs with "M" in the part number are consumer/mainstream type devices using Intel/Micron MLC flash parts and come with a 3 year warranty. The "E" parts numbers are intended for enterprise users and will use Intel/Micron SLC flash parts. Performance characteristics, pricing and warranty have not been published - altough the rumored numbers look very good. The "E" series drives are already sampling, so it won''t be very long before we know all the details. My guess is that Sun will be basing their upcoming SSD products on Intel technology... My reason for posting this to the list is to try to be helpful and let folks know about SSD related news/reviews that they may have missed. There are lots of applications for "M" series SSDs in a ZFS based environment, for example, any application where the workload is mostly read only and the access pattern is random. Serving up CD and DVD sized ISO images comes immediately to mind. In this application a single M series drive providing 4,500 IOPS would be cost/power/cooling/size/price competitive against five 73Gb 15k RPM SAS drives. Regards, -- Al Hopper Logical Approach Inc,Plano,TX al at logical-approach.com Voice: 972.379.2133 Timezone: US CDT OpenSolaris Governing Board (OGB) Member - Apr 2005 to Mar 2007 http://www.opensolaris.org/os/community/ogb/ogb_2005-2007/