Enterprise Flash Drives -- An Update

So, they've been in the market as part of EMC's DMX for a while -- what have we learned?

A lot -- and the story just keeps getting better and better.

Let me bring you up to date on what we're seeing.

A Bit Of Background

EMC surprised a lot of people when we announced EFDs (enterprise flash drives) for the DMX at the beginning of the year.  Industry watchers realized that this was the beginning of a shift to an entirely new storage medium in the enterprise, and had big implications.

Competitors who were caught a bit off-guard told the market some combination of the traditional story: (a) it's not ready for market, or (b) you don't really need it, or (c) we'll have it real soon. 

Some vendors said all three things.

Customers were intrigued on two levels: strategically, what are the implications, and -- more tactically -- where can I use this stuff today, and what's the impact?

The tactical results are in, and you might be a bit surprised.

Yep, They're Faster

If you think in terms of IOPs (input/output operations per second), the headline is that -- running in one of our storage arrays -- they deliver about 30x more IOPs than the fastest magnetic disk drives.

And when we find applications that could really benefit from this level of improved performance (and I mean *really* benefit) the results are astounding.

Some of these applications where they're being used to day are of the more exotic variety: trading systems, mission-critical OLTP environments -- big iron stuff.  But a surprising number are garden-variety: Exchange, customer-facing order systems, and so on.

And we've collected an interesting portfolio of before-and-after stories.  All of them are eye-opening.

An interesting bit of technical trivia involves read caches.  One of the things that sets apart enterprise storage arrays from their mid-tier brethren are large, shared caches that are used to accelerate both reads and writes.

Even in applications that were getting high cache hit rates for reads (e.g. greater than 90%), the overall acceleration in performance was notable as well.  Turns out that when you miss from cache, you're hitting a traditional disk drive, and -- even with this presumably small penalty -- the acceleration on the small number of misses turns out to be a very big number, performance-wise.

Bottom line: almost *every* application can benefit from EFD performance that we've seen.  The question is "how much" and "is it worth it"?  To help, we've improved our tools that can look at a specific application environment, and answer at least the first part of that question.

Yep, They're Cheaper

Enterprise flash?  Cheaper than disk?  Yup.

To get high levels of performance, we've got more than a few customers who buy large numbers of high-performance disks, and short-stroke them to get the very optimum performance.  As in hundreds of drives, in some cases.

A while back, I wrote a post on data placement issues in performance-sensitive environments, and I was very surprised at how many people haven't been exposed to this issue, or don't think it's important.  For all of you that don't think this is important, there's a few large shops I wish you could meet ...

Guess what? 

These customers are justifying signficant quantities of EFD simply on economics alone.  They're cheaper than the alternative.  Much, much cheaper in some circumstances.  Some of the informal quotes from customers are great, e.g. "milliseconds saves millions" (of costs, I presume) and "this was a cheaper alternative than anything else we looked at".

We've gone as far as to create a TCO tool for these situations that look at storage costs, cabinet costs, power, floor space, et. al. to help with this sort of justification.

And, Of Course, They're Very Reliable

I think flash technology in general got a bad rap from the consumer-grade stuff that we're all familiar with.

When we started shipping the technology, we were very clear to differentiate between the different technologies -- there are substantial differences.  Some people didn't get the message, or might have intentionally chosen to ignore it (think vendors).

We were also clear that we had extensively tested these drives, as we do with all drive technologies.  I would venture that no one -- repeat no one -- tests drives as thoroughly as we do.  It's a core part of our business.

Now we can add field run time to the list.  Guess what?  They show ever sign of being as reliable, or more reliable, than the spinning disks they replace.  And these EFDs aren't being put in nice, gentle environment. 

They're getting the c**p hammered out of them. 

That's what they're designed to do.

So, Here We Are

It's been about six months since we've introduced the technology in the marketplace.  We're even more convinced that this is big -- it's a fundamental new technology that shows every sign of crowding out performance-oriented disk drives before too long.

The various naysayers now have to confront hard customer evidence.  The performance levels can really matter to a business.  In some cases, they're cheaper than the disk drives they replace.  And they last a very long time.

I'm sure that, sometime next year, we'll see more storage vendors offering EFDs.  By now, I bet they're realizing it's not as easy as simply plugging in a new device into an existing array.  There's some heavy lifting involved ;-)

Maybe next year?

Comments

[Shibin Zhang]

Chuck,

NetApp invites me to an on-site meeting about storage performance and acceleration. I am doing homework. This is a helpful article.

Thanks,

Shibin