TSA has done a good if slightly skewed write-up on the current positioning of Enterprise Flash disks and what the various vendors are doing with them. And at the moment, they are all using them as faster spindles; as faster replacements to spinning rust. In the same way that I could remove my laptop/pc drive and replace it with a SSD; I can do it in an array at some extortionate price.
I know there’s been a fair amount of tweaking etc to get them into the various arrays but it doesn’t appear to be the proverbial rocket science! So can we have some rocket science now?
We’ve actually looked at leveraging EFDs in our current environment but at the moment as well as the initial expense; the actual retrofitting into the environment is either a lot of work or it means buying a lot more SSD than we would possibly need. The layout in a shared array becomes arcane and possibly more complex than a mere ‘bod can hold in their heads. This does not mean that we won't do it but it's not a magic bullet; it'd probably be a lot easier if we were not looking to implement into an already running business critical environment.
I think if the industry simply looks at these as faster disk spindles; we are missing a huge opportunity to make storage a lot more efficient. We need to understand a lot more about what actually goes on at an application layer; for example, which individual files are hot and need to be stored on faster storage. In a hot LUN (a concept which must die); what percentage of the data is actually hot, is the whole LUN a hotspot or is it a tiny chunk of that LUN? If it's a tiny proportion of that LUN, is it not really wasteful to put the whole LUN onto this expensive resource.
Things like Sun’s ZFS hybrid disk pools might help a lot; we need to see more collaboration between the Operating System/File System and Storage Vendors to get this all to work efficiently and cost effectively.
No, you don’t need to make huge changes to use EFDs but to use them efficiently and not simply give the greedy storage vendors even more money; you might want to think about making huge changes.
Of course, by the time we've fixed the problems, EFDs will probably have come down so far in price that the economic issues will have gone away. So perhaps the solution is for all the vendors to take the hit now and bring the costs of EFDs down to the cost of traditional disk?
Martin –
I suspect the future state you hope for is a widely shared vision, and there are likely to be projects under development today across the industry that will ultimately help maximize the utility and benefits of flash more effectively than the current state.
But in the mean time, flash can still save you money without adding a lot of complexity, especially if you’re currently short-stroking disks or considering a server upgrade to improve performance. One customer spent a half-million on flash drives and avoided over $3M in z10 MIPS upgrades, so the savings can be HUGE!
And FYI – EMC has already brought the prices of its EFD’s down by over 75% since the initial introduction; unfortunately competitor pricing seems to be stuck in the 30-40x an equivalent 15K FC drive price model from the beginning of last year. Making the numbers work out with that sort of whacky pricing will indeed be difficult, but I’m sure competitive pressures will drive them down fairly quickly (once they’ve each recouped their eval/development/qualification costs, I suspect – it is quite expensive to buy enough EFDs to verify reliability, performance and compatibility).
I think you summed up better here what I’ve been trying to say for the last 18 months… it means more from you as you have quite clearly positioned it from an end user perspective.
I’ve never said to TSA we wouldn’t be doing DS8K SSDs, just that its not the only answer – and all my questions about how many in a DMX were more aimed at getting some DMX performance admissions.
(I can’t comment on his post, since I want to prove him wrong – lol)
BarryW – you can run, but you can’t hide:
http://thestorageanarchist.typepad.com/weblog/2009/02/1039-dont-miss-the-amazing-vendor-flash-dance.html#comments
B^)
Agree – conversation needs to move beyond ‘fast spindles’ to being ‘useful capacity’… Far too much manual layout and point in time solutions right now, no real understading of the opex over time to continue such ‘benefits’. And frankly – other than in some niche areas – my bet is that the money for SSD would be better spent elsewhere with easier / larger realisable benefits coming from elsewhere in the value chain. Low latency storage spindles should be the last point of resort in the performance discussion.
Sub-LUN pages that can be tiered over/between spindle types would make SSD more tollerable – where a qty of pages in a LUN are on SSD, some on SAS and some on SATA, allowing us to ‘tune’ the LUN by dynamically & transparently altering the %s of pages on which type.
Agree re the VM/FS changes linked to use of low latency storage is needed – liking the L2Arch project in OpenZFS etc…
But yes I’m getting bored of the ‘my spindle is better than your spindle’ conversation that seems to be focused around SSDs 🙁
Oh and I’m still wondering just how many SSD/EFDs have actually been purchased by customers in ent arrays? And what % of those achieved the benefits they were expecting at the application level without additional investments?
@Martin;
Good points well made.
I spotted this post from you last night while Calvin Zito (HP) and Chuck Hollis (EMC) were ding-donging it out over on my blog on the very same subject; http://blogs.netapp.com/shadeofblue/2009/02/emc-says-ssd-is.html.
Fair point well made Storagebod (Royle Family Christmas 2008)
But surely this is the way it always works…. everybody rushes to support the new technology, at version 1 its “OK” and a few people take it up, then gradually it becomes more and more integrated with te array, and the array in turn becomes more tuned to suit the technology, then eventually we all wake up one morning and realise that under our own noses the landscape has changed but so gradually that we barely noticed. Evolution rather than revolution.
May be Im wrong (would love it if I was) but I cant see EMC and Hitachi overhauling their flagship products during a single array refresh cycle. Frustrating as it is, these things generally take time.
I think your points are great, I have a huge list of questions for HDS if/when Im in a position to question them about EFD in the USPV. I doubt they will answer them all but I will ask them anyway. I want to know how well its integrated etc before I recommend anyone to deploy it. Like you say, you want real value from them rather than just throwing a faster spindle into the mix and hoping the array will make the most of it.
I personally think EMC have done a great job so far and Im jealous as hell of them. I honestly think they have done about as much as we realistically could have expected. Sure Id love them (and Hitachi and the rest) to radically change the way they do things overnight but none of us really expect them to do that.
Speaking from experience, people whos businesses, and lives, live and die with the USPV have been “gradual” on the uptake of Dynamic Provisioning. They wouldnt be about to jump in with two feet if the next version of the USPV was a radical change.
Im excited about EFD as it currently is, and expect EMC to be realistically innovative to try and keep ahead of the rest. I think we’re in the process of personally experiencing (oh how it grinds that Im still watching from the sidelines) a major paradigm shift in storage.
I would really like to hear just how much EMC EFD costs at this point. TSA mentions a 75% reduction in cost – but from what? To what? We’re all talking about angels until we get some numbers.