While I wasn’t in the room to hear the comment directly, IBM’s new storage chief, Ambuj Goyal, is reported by The Register to have stated that his objective was to move transaction storage away from disk to all-flash arrays. Ultimately, he envisions an IBM that sells less storage.
Later in the article, he clarifies that he isn’t even suggesting flash-assisted disk or hybrid arrays, but all-flash only — probably leveraging Texas Memory Systems (recently acquired by Big Blue) RamSans.
I have my doubts about the readiness of Flash for anything like the heavy lifting of big transaction systems, especially given the memory wear problem that vendors choose (not) to deal with by simply adding a ton of additional memory to substitute in whenever a cell fails and a group of cells is marked as bad. One credit card company told me recently that his CC processing systems, which handle over 1 million card swipe transactions per second, would burn out Flash SSDs within a few minutes of installation given the write limits that currently exist in SLC and MLC memories.
Of course, last time I checked, TMS was also building SSDs out of DRAM. DRAM SSDs don’t have the memory wear issues of Flash, but DRAM memory is volatile, while Flash is not. And DRAM rigs tend to be substantially more expensive per GB, not that anyone ever accused IBM of being the low cost leader.
Goyal’s other talking points didn’t raise my hackles. He thinks that storage virtualization (a la San Volume Controller or SVC) should be deployed to enable non-disruptive deployments of storage and applications that use it. I like storage virtualization too.
He also hangs his hat on IBM’s burgeoning Virtual Storage Center, a management console. I have to agree that management is the missing link in efficient storage and I am delighted that IBM is developing yet another tool set for managing storage. But is it a storage service management play or a storage resource management play or both? And does it work with all storage gear, with all IBM storage gear, or only with select IBM rigs? I hope to learn more about this at Edge 2013 to see whether it is ready for prime time.
I like IBM, having taken my earliest training in IT at IBM schools. And I have known some top notch engineers from Big Blue over the last three decades. I am not sure the Omni et Flash thing makes any more sense to me than Omni et Orbis (everything on disk) mantra that the array makers have been preaching for the last 20 years.
I would like to learn how they justify this direction. Or are we all just chasing the goofy folks behind Evil Machine Corp’s XtremeIO announcement.
Guys, if you think going really fast for a really short time is cool, why not buy a top fuel drag racer?
Oh, that’s right. Top fuel dragsters have a tendency to blow up.
Oh well, I’m starting to flash back to Ed “Big Daddy” Roth and his Rat Fink illustrations from my youth. But I have long since set aside such childish things.
{ 2 comments… read them below or add one }
Jon -
Regarding write wear with RAMSAN, at a recent briefing I was informed you could write to the device at 100% of the write rate for 5 years and not wear out the device. Now, if it is 100% writes I guess we could just have a simple circuit to acknowledge the write and never store anything.
Virtual Storage Center (VSC) bundles all SVC software features except compression together with what was TPC Standard Edition, TPC for replication and adds new analytics around storage tiering.
Thanks for the feedback. Not sure what to make of the RAMSAM comment, unless these guys know what my transaction load is. If I am doing a million transactions a second, versus say 100,000, how soon would it take to violate the max writes of MLC memory of 250K per cell? How much spare capacity do they offer to substitute or swap in for burned out cell groups?
As for VSC, that sounds right, but I still don’t know about the implementation model. I will be investigating this.
You must log in to post a comment.