IBM bungs deduplication in FlashSystem arrays

IBM launched three new all-flash array products, including an all-flash DS8888 monolithic array. It says the three products each have a minimum latency of 250 microseconds, and use IBM's proprietary format flash drives, called MicroLatency modules. Big Blue says the MicroLatency technology transfers data within the flash …

[Anonymous Coward]

If I understand correctly, 'minimum latency' is a bit like saying 'best case scenario latency'. It's not a meaningful number unless we know what the real world performance is like, whether this is with auxiliary services running etc. I'm almost excited, but until the last whiff of marketing spin is dissipated, not quite.

[Anonymous Coward]

I used to sell Flash for IBM. 250 microsec is the best case scenario. You are unlikely to get 250 microsec. That is the in the lab with bare metal min latency. When you add software, such as Spectrum Accelerate (the artist formerly known as XIV) in these new arrays, you are going to add latency. Compression adds a lot of latency. It should still be under 1ms in real world conditions though. I wish they would have come out with this sooner. IBM had good Flash hardware with not bad, but ancient SVC software on top (V9000) or they had XIV which had legacy hardware but good software. Neither one was really a big winner anymore on its own. Most of the Flash IBM has been selling has been of the 900 bare metal variety, as it is the fastest Flash storage out there... provided you have some sort of host based storage management software as it is bare metal. Now that they have XIV Flash, more or less what this A9000 is, that might be a winner. EMC doesn't have anything like it. Their DSSD is essentially IBM Flash 900. The big competitors will be the integrated systems, like VCE and the hyper-converged and, of course, the cloud. IBM still has no story around converged or hyper-converged.... maybe some of those cool new sub-system vendors like Tintri too will be competition.

IBM keeps calling this Flash stuff "the cloud". They need to knock that off. It just confuses people. A big box of on prem racked Flash which you pay for up front is the opposite of cloud.

[Nate Amsden]

Looking at the specs for the IBM 900 it doesn't seem on paper to be anything close to DSSD, which on paper is most likely greater than 10-50 times faster. Big part of this of course is PCIe flash, which the IBM Flash 900 does not appear to use(since it only seems rated for 1 million IOPS).

IBM isn't alone here of course, HP has no answer to DSSD at this point either. Though I think the relative market size for a PCIe based flash system is quite small vs the more general purpose SAS flash systems like XtremeIO, 3PAR, IBM FlashSystem, PureStorage etc etc.. Having IOPS in the single digit millions for a single flash array(with 1ms of latency) is going to be enough for just about anyone for a long time to come.

[Anonymous Coward]

Agree, IOPS numbers never impressed me. There are extremely few workloads that can even drive 1m IOPS. It is all about latency. EMC says DSSD can do 100 microsec. I'll believe that when I see it... Agree with your statement on good enough too. Yes, something like a FlashSystem or DSSD will smoke a 3PAR, but do you really care about the difference between 800 microsec and 500 microsec, probably not. At that point, they both have passed the performance requirements. We'll see how long people are even buying storage sub-systems. It seems to be a legacy category.

[Anonymous Coward]

The standard game in performance metrics is to turn off all the optional bits to get the big numbers.

What if you can't turn off the "optional" bits as they're not optional?

A9000 and A9000R are running their code (Spectrum Accelerate) all the time - it isn't optional - you can't dumb it down to make it a FlashSystem 900 (if you want to do that, just buy one of those instead).

Data efficiency also isn't optional - the A9000/A9000R will always be de-duplicating, compressing and pattern removing.

So, feel free to get it tested/validated, but the 300 microsecond / 0.3 millisecond latency being offered cannot be without dedupe and compression and neither can the 250,000 IOPS for a single A9000.

Sure, your mileage will vary, but I don't think you're having your plonka pulled as much as you might fear......

[Anonymous Coward]

Why are we talking about EMC DSSD when it's only offered direct attached and therefore limited use cases? Dell is buying a messy storage portfolio with limited synergies except for a thin management layer.