ludodesign - Fotolia
A year after EMC bought little-known flash startup DSSD and months before it starts shipping anything from the acquisition, EMC president of products Jeremy Burton gave a rundown of DSSD's genesis and its technology.
"Andy Bechtolsheim, he's a pretty smart guy, to say the least," Burton said of DSSD's founder. "He did the original Sun workstation back in the day, and then, most recently, did Arista. His night job while he was at Arista was a company called DSSD. And DSSD is a super high-performance flash product. Is it a software company? No, it's a hardware company. It's hardware engineering at its finest. And Andy's built this insane server-attached flash thing that is going to be 10 to 100 times faster than anything else you've got out there."
We recently talked with Burton about how DSSD will run in-memory databases, and where it fits in EMC's overall flash strategy.
Do you consider DSSD to be a storage system or a server?
Jeremy Burton: We acquired DSSD about a year ago, and people think, 'Oh, it's flash; therefore, it must be a storage array.' No, it's not. It's a block of flash that plugs directly into the CPU. There's no network. It's [PCI Express] PCIe 3.0, and it plugs directly into the CPU. Think server-side flash. Then if you think about an in-memory database, you've got incredible [low] latencies directly to the flash.
I mean, it technically is a storage system. But, by way of a direct connection to the server, it's almost like [direct-attached storage] DAS. Certainly we think the killer use cases are going to be the in-memory databases and the high-performance analytics. You might have a petabyte of data. You're not going to have a petabyte of it in DSSD. The data set you're working on may well be in DSSD, and when it's finished processing, it'll dump it off to the commodity. This is actually the way [high-performance computing] HPC systems work. They call it burst buffer. You suck something in; you process it; you dump it out.
When you offload data, what are you offloading it to?
Burton: Object stores, I think, have only just scratched the surface of the prevalence you're going to see in the enterprise. Last year we acquired a company called Maginatics, which is now a gateway. What Maginatics does, if you're running a traditional storage array -- it could be an EMC, or it could be a non-EMC -- and you want to archive data off, you do it through Maginatics. Underneath Maginatics is an object store. So, it implements a layer on top.
Like DSSD, other companies such as IBM buy NAND flash and make their own flash modules. Is that a growing trend?
Burton: To get the density, you've got to do it. Potentially we can use those flash modules in XtremIO as well. [EMC's] XtremIO [flash storage array] has got all the data services, the snaps, the replication, dedupe, compression. DSSD has none of those.
DSSD doesn't do any of those things?
Burton: Yeah, because its [low] latency is everything. If you add compression, dedupe -- the more things you add in the data path, the more you slow it down. For the [DSSD] use cases, you think these new apps, the way they're architected, that you basically have lots of nodes, and you just scale 'em out.
Is DSSD able to pool flash among different servers, or is the data stored in silos on single servers?
Burton: No, you're going to be able to scale this thing out. One DSSD brick is going to be this year 144 TB, going to 288. So think early next year 288 TB of flash that you can then scale out that plugs directly into the CPU. You could have a petabyte of flash plugged into the CPU running an in-memory database.
What are the primary use cases for DSSD versus your XtremIO flash storage array?
Burton: XtremIO is running VDI and Oracle databases and a lot of the workloads that have been in the data center for many years. If I look at something like DSSD, DSSD is going to run underneath Hadoop and a lot of the next-gen apps because it's not a storage array.
What's the strategy for all-flash XtremIO versus hybrid arrays combining hard disk and solid-state drives?
Burton: There's obviously big hysteria around all-flash arrays, which we're big into as well. But, you're going to find hybrid arrays have got use cases with databases, email servers.
The all-flash array, I always look at it as a better way. It's a better storage array. But, the workloads that it's going to run are the same workloads that have been in the data center for the last 20 years. A lot of the new [workloads are] going to go software-defined on commodity, so things like Hadoop. It's absolutely going to be a different architecture. … There's not one workload. They're all different. We think flash arrays are going to be huge, and we're going to make a load of money there. But we're not falling in the trap [that they're] the be-all and end-all of storage. I mean, people couldn't afford to buy an exabyte flash system.
Do you foresee more economical ways to run flash with TLC or 3-D NAND?
Burton: An archive is a write-once scenario. Maybe you can use TLC flash. But, if you look at the economics versus your cheapest spinning media, it's a 10x advantage [favoring spinning media], and the curves from what we can discern don't meet. But, in certain use cases, where you need the [low] latency and you've got deduplication, you can make the economics work. So a VDI use case with dedupe or a database use case with compression, maybe you get a 5x [data reduction] on the dedupe and a 3x [reduction] with compression, you can bridge that 10x gap sometimes.
Does your database story center on the traditional VNX and VMAX storage arrays or your XtremIO all-flash arrays?
Burton: It depends on what the customer is trying to do. The beauty of having a portfolio is whether the revenue goes to VNX, VMAX, XtremIO -- you think I care? I care that it comes to EMC. Over the years, a lot of people are pointing the finger and saying, 'EMC, we don't understand why you've got a portfolio.' I think now this is where the portfolio benefits because if you go to the customer and say, 'Look, we think a chunk of your workloads are going all-flash. We've got one of those. We think a chunk is going to go software-defined on commodity. We've got one of those.'
Analyst sees EMC's DSSD acquisition as bellwether
VP provides update on EMC's flash strategy
EMC's all-flash options include XtremIO, VNX-F