Hopefully, some of the Xiotech folks have their ears on. In response to my previous post questioning the marketecture around Flash SSD, one reader (Pq65) said that my commentary underscored Xiotech’s limitations in the sense that they (1) are closely tied to Seagate, which doesn’t have a Flash story yet, and (2) were not jumping on the Tier 0 Flash story like all of the other vendors. I will only take a couple of seconds to respond to this comment — mainly in the hopes of getting a conversation started.
First, readers here know that I have concerns about the use of Flash SSD in a heavy transactional environment. While I believe that Flash SSD has great potential, even with its current limitations, to store write once/read many data, as is its use by Isilon, I don’t see it as necessarily viable for normal read/write workload — depending on the workload involved.
I have done some calculations for a couple of my clients that demonstrated, given their workload characteristics (update frequency and data volume), that they would be burning out their Flash SSDs and replacing them frequently if they used them as “disk drives on steroids” as some brand name vendors are recommending. Simple memory wear accounts for this. Even with MLC chips, we are stuck at about 250K writes per cell.
Under enterprise workload, I wonder whether Flash SSD is economical at all. If you want a solid read/write accelerator in memory footprint, I would spend the big bucks on dynamic RAM chip SSD and not Flash. Dennis Martin starts to look at this in his article in TechTarget’s new PDF version of Storage magazine, but unfortunately dwells on caching and short stroke disk to make the same tired straw men arguments about SSD versus disk that we’ve been reading for a long time. (Interesting that, since the title of his article is “Why Now is the Time for Solid State.” Hmm. Of note, several of the ads in this PDF magazine are coming from vendors with Tier 0 Flash SSD marketing prominent.)
We will shortly be reviewing Nimbus Data’s offering in Flash SSD arrays. These guys are willing to put their money where their mouths are and to allow my labs to beat their unit up to get a sense of actual durability and performance over time. Kudos to them.
Now, Xiotech. Under new management, these guys are pushing forward the most innovative disk subsystem architecture I have seen in a long long time. It is noteworthy that EMC wanted it killed in its cradle. Why would this be the case if it didn’t scare them?
With all respect to the hardware engineers at what was formerly Seagate’s Advanced Technology Group (now Xiotech’s Intelligent Storage Element/Emprise branded products), hardware is hardware. They did some interesting things with drives/chassis to get better cooling and RAID-like functionality without RAID and disk spares (see RAGS). Plus, their close relationship with Seagate gives them access to the technology for fixing about 95% of disk problems in place — without cracking the case to replace disks thought to have failed. This is really cool. They can fix drives in place most of the time using the same diagnostic/repair firmware that Seagate would use if you sent them a drive you thought had failed.
But I truly care LESS about the box of disks than about the MUCH BIGGER design value of ISE/Emprise – particularly as it relates to
- Xiotech’s decision to eschew the embedding of “value add” functionality on its controller (preferring to work with an ecosystem of third parties for things that should NEVER be done at the controller level — like de-dupe, thin provisioning, etc.)
- Xiotech’s embrace of open management based on Web Services and REST (I can set up their arrays and parse out their capacity with an iPAD if I want to)
- Their willingness to share their “secret sauce” technology openly with all other vendors who want to similarly enable their own products for common management via open Web Services standards (see cortexdeveloper.com)
- And their decision not to gouge customers with high priced warranty and maintenance agreements (five years free is their motto — mainly because the box is so resilient, they don’t have to do frequent maintenance work or even force you to buy spare drives).
Flash SSD will not be part of the Xiotech technology probably until it is part of Seagate’s technology — that is true. But remember that Seagate already flirted with Flash on a drive intended to save notebook battery power in the face of Microsoft Vista’s log-even-while-hibernating foo a couple of years ago. They hybridized Flash and disk on a drive but, to their dismay, Microsoft did nothing to facilitate the use of the Flash components. A lot of work and cost amounted to little revenue. My understanding is that they are working the Flash thing to come up with workarounds to the endemic problems that currently confront Flash SSD. I would just as soon wait for the bugs to be worked out rather than be one of the early adopters of a technology still overpriced and subreliable for what it portends to offer in terms of speeds and feeds improvements.
I do not fault Xiotech for being a laggard in this trend area that all of the name brand array vendors seem to be pursuing with a vengeance. Seems to me that they have the customer in mind, and not just their own top line growth. I like that in a company.
Flash SSD is the new shiny thing. But, I guarantee that it won’t make you any more popular on date night to brag to the guy or gal you are trying to impress at your local bar that you bought x TB of overpriced Flash SSD. It might actually piss off whoever controls the purse strings at your company if you are trading off a double digit performance bump (probably non linear and non sustainable) for a bump of a couple hundred percentage points over the obscene prices you are already paying for name brand storage.
That’s my take. Tell me why you think I am full of horsefeathers…
{ 1 comment… read it below or add one }
I’m not sure how you did your math to come to the conclusion that SSDswould “burn out…frequently”, but I suspect you’ve not taken the architecture and implementation into consideration. Your math is perhaps made more suspect by your assertion that “MLC chips…are stuck at 250K writes per cell”.
The rated program/erase cycles of MLC are more like 10K, while SLC enjoys ratings of 100K (and are frequently observed well in excess).
If you consider the architecture of the STEC ZeusIOPS family of drives, where the internal logic actively relocates blocks across an internal 50-60% overprovisioned capacity of NAND, the nath shows that you could do 100% writes to these drives 24 hours a day for just under 4.8 years before you will start observing loss in writable capacity – at the RATED 100K P/E cycles. And this holds true whether you overwrite a single block for that entire life, or randomly distributed writes across the drive’s LBA range.
Since most applications are not 100% write, these drives well exceed a 5-year projected lifetime for any application…and the device’s lifetime can be further extended by intelligent alignment and buffering of writes, making the issue even more moot.
Coupled with logic that distributes the LBAs of a LUN across different classes of drives (SSD, FC and SATA) based on performance and workload, customer can attain much higher performance at lower response times than with pure disk solutions, and at a cost equal or LESS THAN the pure-disk approach – yes, even at today’s economics.
You must log in to post a comment.