BOSTON -- At Dell Storage Forum earlier this month, Carter George, the vendor’s executive director of storage strategy, outlined plans for Dell’s flash-based Fluid Cache technology roadmap.
In an interview with SearchSolidStateStorage.com, George explained the Dell SSD strategy, including the motivation behind the server-based Fluid Cache PCI Express (PCIe) card as well as a Fluid Cache appliance, which he said would emerge late in the first quarter or second quarter of next year.
George also supplied details about another Dell SSD project, code-named “Big Iron,” that could lead to an all-flash array in Dell’s Compellentproduct line. George said the all-flash array is currently “under investigation.”
Do you view Fluid Cache as Dell’s answer to what EMC has been doing with server-side flash?
Carter George: Yes, part of it. I mean, EMC’s doing a lot of stuff. EMC has so far announced three flash things: [Project] Lightning, or VFCache, a host cache; [Project] Thunder, a cache appliance; then, XtremIO, which is a flash array, and that’s not shipping, either. I think people will find you need to do all three. The only one that could be optional is the [cache appliance], but I actually think there are good reasons to do it. An appliance is a great way that you could go stick a bunch of flash in the network and use it as a cache without having to touch your servers.
This is a really big deal, the transition from disk to flash. In three years, all active data is going to be on flash, or some kind of memory-based storage, and that means fundamentally new array designs. It’s not just sticking SSDs in your existing disk array. It just doesn’t work.
I think EMC correctly saw this as a legitimate threat to their franchise with Symmetrix.
When we first started talking about this, we were thinking, ‘Hey, here’s the opportunity to go knock EMC off their throne.’ But, they saw it just as clearly as we did, and they’ve been investing like crazy. They paid almost $500 million for a company with not a single dollar of revenue [XtremIO], and that tells you that they’re not leaving anything to doubt in this space.
So, if the opportunity isn’t there for us to destroy EMC, too bad. But, I do think there’s an opportunity for Dell to be the other company that gets this right. I think Dell is well ahead of everybody else but EMC in this flash space.
George: We have two things: Fluid Cache and an appliance that, for now, let’s call the Fluid Cache appliance. Thunder and Lightning are both read-only caches, and ours are read and write. That seems like a simple difference, but it makes all the world of difference. There are two things you do in storage: read and write. They’re only doing one, and we do both.
To get things like snapshots and replication to work correctly, if you’re a read cache, those things are easy. But, if you have a write cache, you have dirty data -- new data sitting up here in the cache -- that’s not down [in the storage array]. If I take a snapshot, that’s a point-in-time consistent image of my data that needs to include that dirty data. We’ve done all this coordination to have snapshots be aware of where the dirty data is and flush it or include it, one way or the other.
Same thing with replication. If I’m replicating Compellent to somewhere else, it’s going to replicate not only what’s here [in the array], it’s going to replicate what’s up there [in the cache], because it knows about it. That’s where all the hard work is, coordinating all those things.
What else has Dell done to address challenges with server-side flash?
George: To have it in the server, it has to be reasonable for you to be able to put the flash in there. If you’re buying a new server, great. Put it in the server. It’s the right place for it to be. It’s on the same bus as the CPU. Access times are better.
But, if you have an existing server that’s in production, putting PCIe cards in is a pain. You have to bring the server out of production, take it out of the rack, put a card in and hope that all works out, and then put it back in. People don’t do that with production servers.
We did a bunch of things just a couple months ago. We launched 12th generation servers, and these servers have a bunch of stuff in there specifically for this. Dell and Intel cooperated on a new standard called NVMe, non-volatile memory express. This both defines a software driver standard and a physical connection.
In a server, you have front-facing drive slots. The NVM connector lets you put something in that drive bay that is hot pluggable like a disk drive but is actually a PCI device. It sits directly in the PCI bus. If you have one of these servers, you can go in and just stick one of these [NVM devices] in the front of the server.
Will solid-state cache be the main Dell SSD play?
George: Yeah. These cards do 250,000 IOPS each. That’s really fast. That’s about as fast as a Compellent [array]. And I can have four of those in a single server, so that’s a million IOPS. That’s faster than our biggest storage array. I could have 256 servers with a million IOPS of cache each. So, that’s hundreds of millions of potential IOPS.
This doesn’t get you out of wanting to have a flash array. It gets you into it, because if I start going really fast, and I’ve got a bunch of dirty data up here in the cache, and I had a snapshot point or a replication point, this cache wants to be able to flush down here [to the array] to get consistent. And, if there’s a huge difference between how fast your cache goes and how fast your backend storage goes, you’re going to have weird performance, where you go really fast for 30 seconds and then sort of slow for 10 seconds, and then really fast for a minute, and then slow for 20 seconds.
People aren’t going to want that. They’re going to want consistent, even performance. So, as this cache goes faster and faster, you’re going to want this [array] to keep up. That’s why I said you need to have all three pieces, [host cache, a cache appliance and a flash array].
Do you think Dell needs to make an acquisition to achieve its flash vision, or can the company do the work internally?
George: We can do it internally, and it would take about two years. There’s a project called Big Iron to design an all-flash, scale-out Compellent using [technology from Dell acquisition] RNA [Networks] in the Compellent head to help with the scaling out and make changes in the code, optimize for flash, and do the wear-leveling and all those things you need to do, the in-band dedupe, not the post-process dedupe. Four or five pretty fundamental things we’d have to do to Compellent to make it an all-flash array. By the end of this year, we will have to make the decision if we can wait two years. Is two years too late? Two years might be too late. If we thought it would be the best thing, but the market will have been settled by the time it happens, we’re going to have to do something else.