In the industry press, there's a certain group of people who disdain the enterprise approach to storage, thinking the world's important information will be entirely stored on commodity boxes running open source software.
Nothing could be farther from the truth.
High-end storage arrays such as the DMX reflect the very latest in storage technology. And there are a lot of very serious IT people out there who won't run their business on anything else.
Today, EMC announced some fascinating upgrades to the DMX -- including one that I'm sure will absolutely intrigue most storage techies like myself -- enterprise flash drives.
But, as with most things that EMC announces, there's two stories here: not only what we did, but how we did it.
The IOPs Conundrum
Most casual users of storage focus mostly on cost for capacity. They don't usually give a whole lot of thought to the other storage metric, and that's IOPs (input / output per second). An IOP is basically one read or one write.
If you're running a classic transaction processing system, or an advanced web application, you care about IOPs way more than you care about capacity.
Why? Every transaction, or IOP, represents money. If you're booking airline seats, or shipping iPhones for the holidays, or making arbitrage trades in the currency markets, there's a direct correlation between IOPs and money.
If you're familiar with these sorts of applications, you know exactly what I'm talking about.
As disk drives have gotten bigger, they haven't gotten that much faster from an IOP perspective. So, if you needed more IOP performance, you usually had to lash a large number of disk drives together, stripe and short-stroke them, and end up wasting a bunch of capacity / space / power in the process.
And, even then, that approach was limited. Hard disk drives are far slower than other components in the system, e.g. processor, memory. We (and our customers) wanted something far better.
The Enterprise Flash Drive
Enter flash memory technology. EMC announced a purpose-built enterprise flash drive that, externally, fits into the normal disk carriers of a DMX. From most aspects, it just looks like a disk drive, except for one important attribute: it delivers the about the same number of IOPs as roughly 30 standard 15K 300GB drives.
Not only that, from a customer application perspective, each IO has 1/10th the response time as any single or pooled disk drive can do.
Now, many people will say "gee, it's so much more expensive than disk", and -- on a capacity basis, they're right.
But shift your measurement system to IOPs, and you'll notice that flash disk drives can be far more attractive than other approaches.
Oh yes, and then there's the power angle as well. If you look at it from an "energy per IO" perspective, this flash drive consumes about 98% less power than rotating disks.
And, if you own a DMX4, these pretty much just plug in alongside the other disk drives already in the box. No drama, no fuss. That's cool.
This Wasn't Easy
Lest you think we simply took a trip down the street to our local components distributor, and just slapped this in, let me tell you there was a ton of work behind this.
First, this flash technology isn't the consumer-oriented SSD stuff you might be familiar with. That sort of flash technology can wear out, and is optimized for sequential (bandwidth) performance, rather than fast response times.
We're using an entirely different flash technology (Single Layer Cell NAND flash-based persistent storage), and then had to work with the vendor to make sure all the non-trivial stuff, (e.g. dual-ported FC interfaces, the DDR cache, on-board power backup for destaging, the multi-channel parallel architecture to improve performance, and even the packaging) was up to the rigorous standards of “enterprise class”.
The result is essentially a new class of storage device: the enterprise flash drive.
Not easy stuff, I doubt you'll find it anywhere else anytime soon.
And there was more work to do on the DMX: the Enginuity (microcode) algorithms had to be seriously tweaked to understand the optimum performance characteristics of the new device, block alignment, pre-fetch, etc. And, of course, an extended set of tests to make sure these new enteprise flash drives could be used alongside existing disk drives.
And, most importantly, the work to make sure that the entire extended feature set of the DMX "just worked" with enterprise flash drives.
I don't think you'll see that anytime soon from anyone else, either.
And The Result?
I mentioned before that "how" EMC did it was just as important as "what" we did. Because it looks just like a disk drive, just about every feature and capability of the DMX now works with flash drives.
Mixing and matching this new "tier 0" of storage with all sorts of other disk drives. The ability to perform 100% non-disruptive upgrades and reconfigurations. The most advanced local and remote replication available. Priority controls and cache partitioning for different workloads. Security enhancements. Etc. Etc. Etc.
Everything just works, except it's much faster.
From my perspective, that's extremely cool. EMC did the work so customers can just plug it in and go ... go faster, that is.
More Tiers In A Single Array
You may not have a part of your environment that's crying out for more transactional performance or incredibly short response times -- that's the case in many shops.
So there's something in the EMC goodie bag as well for you -- the availability of the new 1TB disk drives you've been hearing about. We skipped the 750GB drive and went right to the 1TB drive.
Like the flash drives above, they plug in without drama, and everything just works.
Most of our DMX customers have different tiers of storage in the same array, now they can create higher highs, and lower lows -- if they need.
It's funny, I read press releases from some of the other smaller storage vendors claiming that their approach to striping or clustering or whatever eliminates the need for different tiers of storage. Forgive me for being skeptical, but I'd be plain amazed if they've figured out how to span the range between flash drives and 1TB drives in terms of performance and cost tradeoffs.
But Wait, There's More ...
You'll also notice in the press release the availability of Virtual Provisioning for the DMX -- a feature where you can provision virtual volumes to applications, and consume physical storage behind these as applications need it, rather than ahead of time.
Now, the skeptics among you will say "gee, that's nothing more than thin provisioning, that's been around for a while", and -- up to a point -- you're right -- thin provisioning has been around on other storage products (including some of EMC's) for a while.
But this is different enough to warrant a closer look, I think.
First, when you dig into the implementation details around most thin provisioning features, you'll notice that there are a bunch of restrictions, e.g. can't be used for performance-sensitive applications, doesn't work with local or remote replication, and so on.
Worse yet, many of the implementations don't report out correctly to storage management tools, creating all sorts of potential havoc when you have one view of capacity that applications see, and a very different view of capacity that the storage array sees.
What you'll see with the DMX implementation is a relative lack of what I call "marketing asterisks", you know, the fine print where they list out all the caveats and things that don't work.
That means that virtual provisioning can be used very widely in almost all situations you're likely to encounter on a DMX. Sure, it can save money on capacity, but what it really saves is time -- administrators spend far less time provisioning and configuring storage, and -- more importantly -- not worrying too much about what works with what.
And, of course, virtual provisioning works with the new flash drives (even more useful here, if you think about it), as well as the new 1TB drives.
Flash-Is-To-Disk As Disk-Is-To-Tape
So, I'm usually asked by customers about how to think about these technologies in the long term.
As an example, despite cries of "tape is dead" by the more extreme among us, tape is around, and will be for a very long time -- we'll just use less of it because the alternative (disk) is becoming more attractive.
I think the same is true of enterprise flash drives: disk isn't going away anytime soon, but we'll use less of it as the newer alternative becomes more attractive.
But, keep in mind, as this new technology becomes more popular, we'll start re-thinking other aspects of storage design, everything from protection schemes (e.g. RAID) to the new role of cache to I/O interfaces. One thing can lead to another.
I see this as the seminal start of a complete re-thinking of storage devices and the arrays that deliver them.
But Where's The Real Story?
Taken together, this announcement is more than your usual we've-got-a-new-product press release.
EMC has developed an entirely new class of storage device -- the enteprise flash drive. We made it work transparently in the single most successful enterprise storage array in history -- the Symmetrix.
And we delivered a feature that helps you use it very intelligently -- virtual provisioning.
I know that there will be the usual skeptical and/or cynical responses from those out there who don't particularly like EMC, but it's turning out none of that really matters much in this case, since we already have a significant number of customers who've agreed to deploy these drives. Based on what we've shown them, they want a bunch of these new enterprise flash drives. As soon as possible, please.
But let's step back even further.
I don't know if you noticed, but EMC is on a bit of a roll in delivering innovation in the storage marketplace. In quick succession, we've seen a new home storage product (EMC LifeLine), an entirely new SMB VMware-oriented entry array (AX4), and now this.
That'd be enough to talk about, but -- I'm just itching to tell you about the rest of it! I guess that'll have to wait for another blog post ;-)
[for more on this announcement, please go see the extended announcement materials at EMC.com, or visit The Storage Anarchist and StorageZilla for their takes]
[Other links include Chris Mellor over at Techworld, the WSJ, and the obligatory press releases here and here. If you're really interested, check out the spiffy multimedia press release here.]
awesome chuck! although these have been around for awhile i've been waiting for a big, quality name to start driving implementation!
would you mind letting us know whose flash memory/controller EMC will be using? is it sandisk, micron, STEC, samsung etc...?
if you dont tell me you'll force me to go buy one and tear it down! ;)
Posted by: shaun noll | January 14, 2008 at 10:41 AM
The 30X IOPS increase is fantastic.
Please tell me that the price per GB is not also 30X.
What's the MTBF ? On one hand, I'd like to think that a solid state device is going to be more reliable than an semi-mechanical fast-spinning piece of brown stuff. On the other hand, filament lightbulbs aren't exactly built to last.
Posted by: mgbrit | January 14, 2008 at 01:14 PM
The press release said it was STEC, they issued a press release this morning. I don't think you'll find the specific EMC build generically available, though.
Posted by: Chuck Hollis | January 14, 2008 at 05:03 PM
I'm not sure of the exact pricing, but I'm sure it won't be exactly cheap, either. I am confident that the costs will decline dramatically with volume, sooner than most might expect.
As far as MTBF, again, I need to do some research, and I'll come back to you.
Posted by: Chuck Hollis | January 14, 2008 at 05:06 PM
I'm interested that you only claim 30x performance. One of these drives will sustain 50,000 read IO/s and 19,000 write IO/s.
Given that a standard drive will do 350 if you are lucky, I make that 142x on reads, and 54x on writes. OK thats at the specific 4K range, and as the transfer size increases, the IO/s reduces. Maybe the 30x is the impact that DMX has on reducing their capability?
STEC provide a 5 year guarantee on these.
As per a previous post on my blog, I've been playing around with these for some time.
However my question is more, how many of these can you actually sustain in a DMX? 10, 15? Thats a pretty big datacentrer footprint for just over 1TB of capacity.
Anyway, be interested to know who many you can support (not how many you will sell to customers) per DMX.
Posted by: Barry Whyte | January 15, 2008 at 05:15 AM