This article is more than 1 year old
Can all-flash arrays, spinning disks and hybrid really live together?
AFA without tears
The ideal: pooling
This siloed approach represents a backwards step in many ways. In the early days of storage, silos were an accepted evil. Storage was typically server-attached, and dedicated to a server-specific set of applications. It meant that the servers could talk quickly to the storage, but incurred the inevitable inefficiencies and overheads, as a lot of capacity went unused.
The move to SAN-based storage enabled IT departments to pool their relatively slow disk-based storage, eliminating the silo effect and increasing usage efficiency. It came at a cost, though, as even high-speed network interfaces introduced latency into the equation, but it didn’t matter as much for pure spinning disk because the mechanics of the drive were also relatively slow, so the network latency didn’t matter as much. Now, SSDs are forcing storage into direct-attached models again, warns Handy.
“SAN usually has a latency of around 30ms, which when you have disks wasn’t that big of a deal, but once flash started being put into SANS, a lot of flash speed was squandered by the SAN,” he said.
So, SSDs leave companies in limbo a little as they figure out how to best integrate them into their infrastructure without leaving too much efficiency behind, warned Alex McDonald, director and vice chair at SNIA Europe.
“There will be a period where we work out how we best use SSDs,” he said, adding that part of this will involve going back to the ‘bad old days’, in which data silos were common. “I'll isolate AFAs in one corner doing these applications and hybrid in another corner doing a different set, and not really being able to pool my storage resource.”
Pooling via private cloud
There is a case for using all-flash and pure spinning disk together in larger environments, say experts. IT teams are more likely to create their own tiered storage using a mixture of the two. The downside for them is that there’s no controller to help them manage the tiering. The answer lies in software-defined storage, Handy said.
“Those are programs that do all of your data management,” he said, explaining that SDS extends into cloud-based storage too. Some companies – he mentions Citi as an example – are moving toward a future where all the slower data is shuffled off to a public cloud provider, leaving only AFAs in the local data centre.
The SDS layer would govern the high-latency data stored in the cloud, and on local spinning disk, and on all-flash arrays (notice how he doesn’t make room for hybrid SSDs here, because they have no place in this picture other than as a siloed option supporting a single app somewhere, and not connected to the pool).
“The software takes over and manages it, so it appears as though it’s all one piece of storage, and appears to be only slightly slower than the flash drives would be if you have an all-flash system,” he says. It’s a lovely picture, but most companies are nowhere near this level of sophistication, warned McDonald, and even if they were, it may not make commercial sense for a few years.
“It’s early days for that kind of cloud orchestration. Very early,” he said, adding that there are still many applications out there that aren’t cloud-native. “We need to let all the old applications die,” he said. “If you’re running 15 year-old applications, then there’s not much commercial sense in moving it from one silo to the cloud, unless of course you can see some cost advantage.”
So, there you have it. Companies are typically either hybrid or AFA shops, or running a combination of AFA and spinning disk if they’re large enough and have the private cloud expertise.
Ultimately, your transition path will depend on your tolerance for operational complexity, said 451’s Robinson.
“One way to address that is by simplifying the number of technologies that you have,” he says, positing a two-tier ‘flash and stash’ architecture as the best possible solution. “You have all your performance-sensitive data on a flash tier, and everything else on a scale-out object tier.”
Whether or not you do that with hybrid or with an AFA/JBOD combination depends on your size and your sophistication. But the general consensus seems to be that you shouldn’t mix the two unless absolutely necessary. Simplicity is key. ®