IBM's FlashSystem looks flashy enough, but peek under the hood... This week IBM announced three new flash products, two of which are based on existing technology. Mainframe/Power customers got the all-flash DS8888, with the A9000 and A9000R models covering the rest of the market. What’s interesting about these latter two products is that they are based on Spectrum Accelerate, otherwise …
No, picture is clearly different from XIV. I've seen Gen 2 up close. Lots of cables on the backside. Power cables on the left, fibre/ethernet cables on the right, all neatly arranged and colour-coded.
This picture looks like a development machine, definitely not a serial production.
So if I've got this right, there's a new DS8000 and a new XIV. They're the same as the old ones but with lots of flash in. They'll perform way better than they did previously but not as well as Flashsystem.
So what's happening with the Flashsystem with SVC bolted on? Are they going to stick any other storage boxes in front of flashsystem? Are there any others left?
Not bolted on. Integrated.
They will continue the high-performance they always did, as well as virtualise all that underutilised storage from third-party vendors, but not in environments where deduplication is the primary consideration.
>Not bolted on. Integrated.
Nope. Bolted on. Flashsystem V9000 is a Flashsystem with an SVC plugged into the top. Yes, it's directly connected via FC rather than through a switch, but it's still two separate products plugged in together.
V700 Unified is a couple of Linux servers sitting on top of a V7000 (which does have SVC code integrated, to be fair).
VSC is an SVC with TPC (itself several products bolted together) and Flashcopy Manager (two completely distinct and unrelated products, depending on whether you're deploying on Windows or *nix)
Of course, doing this may produce a perfectly viable product: Tivoli Storage Resource Manager and Tivoli SAN Manager were bolted together to form TPC and two decades on it's almost usable.
There's nothing wrong with doing it this way: selling Flashsystem with SVC is a good way of getting features without destroying the underlying Flashsystem product, albeit at the unavoidable expense of latency.
Calling it integrated, as with many other instances, a few examples of which I have given, is lazy marketing bollocks. As IBM continues to lose market share it needs loyal customers and duping them with spurious claims isn't going to help that. Flashsystem is probably the best AFA in the market right now, depending on your priorities of course.
Yeah, I know with V9000 it is basically a two node SVC head with a Flash 900 stuck to it. It isn't quite a rudimentary as the V840 where they literally bolted a SVC DH8 onto a Flash 840. If anyone tries to tell you this is special SVC though, that isn't true.
Excellent article thanks, and good work on picking up 'minimum latency'. When I read the release, I spent a while Googling the specs to try and understand why minimum latency was quoted rather than average, maximum or a range. I found *nothing*. Many sites had picked up the press release and quoted it verbatim, going so far as to extol its specified latency as somehow impressive. None seemed willing or able to approach the usefulness of the claim.
After all, my car achieves absolutely brilliant mileage under certain conditions (downhill, tailwind..)
Codysydney,
I don't believe that IBM have done anything other vendors haven't - if anything, they at least did a mixed read/write workload (some vendors just pick read or write) and given that you can't switch off the data reduction capability on A9000/A9000R, the IOPS have to be with data reduction enabled, which does kind of make them more impressive (at least to me).
In these days of data reduction, if you want a fully comparable and relevant set of performance figures to contrast solutions from vendors, the detail required would have to include working set size, compressibility/dedupe result on data being tested, temporal locality of the workload, whether the array had been fully pre-conditioned, whether all data services are in use, etc, etc, etc.
Feel free to add other factors that would be required for comparison purposes that normal punters (including most storage buyers) don't understand.
Don't see anyone being that open unless we get an updated SPC-1 type benchmark (and even that was gamed based on the working set size and using less than the full capacity of the array).
Exactly, its quite the opposite. Everyone of the existing enterprise storage vendors; NetApp, HP, IBM, EMC have taken existing systems, enhanced them for flash with added features and performance and are currently going full speed ahead with great results. :)
We can argue who has done a better job at it but I'll leave that for another day.
Cheers,
Daniel
Agree, that is the industry standard. Take out the disks, put in SSDs, done. If anything IBM is slightly different in that they did not just take an existing array and flip out disks for SSDs... the XIV and SVC software that these are based on has been around for many years.
And I recall that Oracle used to offer the F5100 flash array which was a complete clean slate approach to flash storage (and in my opinion, quite impressive), but it apparently didn't sell well enough to warrant a successor (FS1 is a specific configuration of ZFS unified storage and can be configured as AFA or as mixed storage).
Otherwise, he would have known that the picture he added and even has much to say about is of a development system and not a production one.
When you sit behind the desk for too many years without too much of a practice you show a picture like that and present it as a production cable layout.
Solid write up, but only got half the "under the covers" story...
The killer features in the completely new UI - the advanced search and simplified management that scales to 100 systems. See video here: https://m.youtube.com/channel/UCp75FYeslSfd0ePfQPRFIvA
"Skewed to controllers" - they're all active and by dispersing IO, means that the A9000 can sustain the loss of a controller & the A9000R can sustain the loss of multiple controllers.
Common software (Spectrum Accelrate) - having common software across FlashSystems, cloud services for spectrum accelerate and build your own hyper converged system/SAN provides customers with choice, flexibility and ease of management (with common features - QoS, multi-tenancy ect) .
Full disclosure I work for IBM. Views are my own.
I always thought that this IBM software defined storage strategy was a miss. I get it - software = high margins, hardware = low margins, IBM just wants to sell the software, but it is just confusing. IBM defines SDS as taking existing sub-system software and applying it elsewhere. Everyone else defines software defined storage as having the hypervisor manage storage. So when people hear SDS they think IBM has something like VSAN and are confused. I guess if VMware can call ESX "the cloud" (even though it is the opposite of the cloud), then IBM can call XIV and SVC software defined though.... the first is more of a reach than the second.
"They think ibm has something like VSAN and they're confused"
It's one of those best kept secrets, IBMs Spectrum Accelerate, the same software that runs the A9000, also runs the same way VSAN does - hyperconverged on x86 servers - with full VMware support/integration.
Forgetting the little details KVM support etc, the major difference between Spectrum Accelerate and VSAN is, that with VSAN it's a license that only ever lets you run a hyperconverged environment. Spectrum Accelerate gives you the flexibility to build your own SAN or deploy it as hyperconverged - or do both but manage the resources from a single interface.
IBM is lagging behind on relaxing the limits on its supported configurations and definitely doesn't have the market awareness that VSAN does, but with its latest software stack it's definitely gotten more interesting.
Yeah, but IBM means you can use the publically available VMware APIs, as can every other storage system in the world. That = VMware "integration", it isn't natively integrated. Put another way, how is what Accelerate or XIV can do any different than what any other storage provider with VMware API support can do? The whole idea of converged or hyper-converged is that it is... converged... and all designed to natively function as one. IBM is basically saying you can integrate our storage array with a server. I know, but if being able to integrate, after the fact, is "converged" then there isn't a system in the world which isn't "converged."
Ah the good old "what is hyperconverged" discussion. I agree that many vendors have gone way beyond simple VCentre integrations, Nutanix for example are aiming to get rid of VMware as your management layer all together.
I'd still suggest that the IBM SDS story is better than most traditional storage vendors. For example Spectrum Accelerate is similar in many ways to EMCs ScaleIO.
ScaleIO is fundamentally an island that only runs on servers (hyperconverged or as a build your own SAN), but it doesn't natively integrate with any other EMC SANs.
Where as IBMs Spectrum Accelerate can be transferred from x86 servers across the new A9000s & XIV appliances and also can be consumed as a managed SaaS offering. Full license portability & single management console (Hyperscale manager).
So I guess it really depends what you're looking for from SDS, I think maximum flexibility, license portability and simplicity across deployment models is a pretty good start.
EMC is a bit unusual in that they don't want to duplicate the functionality of their sister companies, VMware and VCE, so they have intentionally stayed in their legacy sub system attached lane with the EMC label while cleaning up with VCE and VMware. IBM has a comparable strategy to EMC... they just don't own a VCE or VMware SDS to compliment that strategy. EMC doesn't care if VCE (converged and hyper-converged) and VMware wipes out EMC, probably hope it does in an ideal situation. IBM needs a VCE and a VMware if they want to stay relevant. HDS has the same issue, but at least they do converged now.... I think IBM doesn't really care though. Their bet is that the lion's share of the infrastructure market will be in a public cloud in short order so why spend time and money working on this converged hardware intermediate step, just go straight to public cloud. The issue for IBM with that strategy, even if they are right (probably are right, just a matter of time) and someone having hardware on prem is just odd in 5 years, is that Azure and AWS dominate that space.
First, when we're asking the servers/controllers to perform all copy services, dedup, compression, etc... then it makes sense the hardware will be skewed to controllers. They're doing all the work!
What's wrong with using existing software and experienced developers to make a new product? I see several improvements over XIV besides dedup and flash media. EMC had to buy XtremIO as they certainly weren't going to re-use the VNX architecture. Pure and other start-ups had to write new code for obvious reasons.
XIV I/O performance wasn't limited by the backplane, it was clearly limited by it's use of SATA drives. XIV was marketed as high-density, grid scale, easy to manage with IBM white glove support - who cares about cabling.
Maybe we need to adopt an industry standard on reporting latency because they're all misleading if you don't know the details.
Finally, let's stop getting excited about vendors reporting millions of IOPS and especially when they're not using realistic workloads or small block cache hit. There are very few companies in the world who'd drive 1M+ IOPS to a single storage array.
Agree with your comments on XIVs controller centric architecture. Not sure why anyone would argue against that, every cloud service runs a "controller centric" architecture. That's how stuff works now. True that latency is the only thing that matters for 99% of workloads which are not going to push 10m or 1m or 200,000 IOPS. Agree that the bottleneck on XIV was definitely the drives and not the backplane (obviously) so there is plenty more performance IBM can get out of XIV if they add Flash.
I disagree that IBM reusing their legacy stacks isn't a problem though. XIV or Accelerate is better than SVC, but if you compare both to say Tintri or VSAN... you are going to see some stark contrasts. They just eliminate the entire, unnecessary, storage abstraction layer and add all of the up stack monitoring. It basically eliminates storage admin and makes everything simple. IBM needs to create or buy a Tintri-like software stack.
This post has been deleted by its author
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
