Opinion

Missing the flash point

posted on 30 June 2008 12:38

Consolidate flash SSD storage to share it

This is a personal opinion expressed by Barry A. Burke, Chief Strategy Officer, Symmetrix Product Group, EMC Storage. These are his personal observations and opinions and do not necessarily reflect the opinions of his employer, EMC Corporation.

Friday again, and I’m perusing Blocks and Files. Amazing how all the discussions about Flash SSDs can be so contradictory.

The folks at Infortrend are clearly not dealing with flash drives that are in the same class as STEC’s ZeusIOPS, which can (and do) handle far in excess of 200 write IOPS per drive – more like 5-7,000 8KB writes/second, and upwards of 15-17,000 512-byte write IOPS. Infortrend apparently are speaking of the much slower drives that are on the market (or just now coming out), yet they talk as if all flash drives are too slow for practical use.

This is clearly not the case, yet from this article readers could be easily mislead.

Then there’s Mr. Jones from Sun, arguing that Flash doesn’t belong in external storage because the CPU in HIS external storage device (Thumper) gets IOPS-saturated by only six SSDs. Fact is, if 6 flash drives overload the CPU in a Thumper, then so would the IOPS workload that 150- 180 15K rpm disk drives can deliver. I’m not sure if Thumper even supports that many disk drives, but Mr. Jones’ assertion is more practically describing the limitations of his company’s external storage offering than it is the upside of Flash drives moved closer to the server’s CPU. Not to mention that the server’s CPU would be just as overwhelmed as Thumper’s, since they could in fact be the same type of CPU!

And I’m sure Sun’s story will change once Hitachi adds flash drive support to the USP-V that Sun resells (as the 9990-V).

But the thing that Sun (and others) seem to be missing in all of these discussions is that you don’t really have to get all of the possible IOPS out of a flash drive for it to have real benefit and value. In fact, environmental concerns aside, if IOPS were the primary buying criteria for storage, dozens of 15K rpm drives will do the job just fine (if your server and/or storage can handle the workload, that is). But customers have been going that route for years, and it’s just not enough.

What flash drives really deliver is faster IO – lower response times than are possible from ANY disk drive (yet still slow than the response time you can get from a DRAM cache).

All this talk about installing flash drives in servers seems perhaps to be an attempt by server vendors to recapture the revenues they’ve been cut out of by external storage consolidation. But the fact is, Flash inside the server won’t beat the response times of a good cached disk array (like Symm, or USP-V)…these arrays today routinely deliver cache-hit response times that Flash drives can’t match – no matter WHERE they are in the infrastructure. Fast cache in front of (relatively) slow disk will consistently beat an uncached flash drive on both IOPS and response times (if the cache is large enough for the workload). And that’s including any overhead that the SAN imposes, which really isn’t very significant, despite what Michael Workman, the CEO of Pillar Data Systems would have us believe.

And sure, servers could get a little better performance if the flash drives where embedded – but the same is true for DRAM, yet it is often more cost-effective to centrally consolidate the RAM in the storage array than it is to distribute across multiple servers.

And given that Flash is still relatively expensive, many customers will not be able to justify installing a mirrored pair of flash drives in every server, much less whole a RAID 5 group of them…this, it is far more cost-effective to consolidate all the application workloads around the data center on a rank of shared flash drives installed in a central storage platform, allocating a portion of the total flash capacity to the different applications. This is the same logic that justifies half-terabyte SDRAM caches in storage arrays – the expensive resource is more cost-effectively acquired in bulk, and it is more performance-efficient and cost effective if they are shared across multiple servers and applications.

And the crazy thing is, once the storage gets put back into the server as all these vendors are suggesting, the issues of backups, replication, business continuity and disaster recovery raise their ugly head – not to mention data sharing, migration, and application development – it’s not very easy to take a snapshot of the production database to run the next release of an application against on the development servers if the storage is captive inside the production servers, for example.

External storage consolidation has a defined and established value, and Flash Drives enhance that value rather than give cause to revert back to the days of embedded storage.

I’m pretty sure you understand all this, but the casual reader of traditional press OR of your Blocks and Files will undoubtedly be mislead to believe that

a) Flash is too slow for write applications

b) Flash is so fast it needs to be in the servers instead of central storage

c) Flash is too expensive to be practical

d) Centralized storage would limit the performance of flash to the point of uselessness

e) Flash is too unreliable to be practical

f) All flash drives are effectively the same

None if these are true…and many of the vendors taking a contrary position today will be contradicting themselves as flash drives gain broader adoption and traction.

Market acceptance and demand for Flash is being hampered by the confusing messages being presented by the various vendors and spokespersons. Whether the misinformation is born out of naiveté for the technology, is in defense of their own delays in adopting and integrating the technology, or is simply the old FUD-machine trying to slow things down so that they can catch up, it doesn’t really matter…the result contradictory positioning is slowing the adoption and hence the cost erosion that is dependent upon volume.

Because the NAND chip manufacturers aren’t going to cut component prices and increase production until the demand is sufficiently high enough for them to make their profits on the volume…until then, they are happy premium pricing SLC technology while Apple iPods consume the MLC volume.

IMHO, that is ;*)