SFD5 Prep-work: PernixData

PernixData architecture

Source: PernixData

PernixData sell storage cache software, essentially.

Their Flash Virtualization Platform, or FVP, is a software shim that sits between the Hypervisor (VMware only, according to their datasheet [PDF]) and the SAN (block only, so FC, FCoE or iSCSI attached, no NFS or SMB). It uses server-side flash to hold the cache, using any flash product supported by VMware.

The selling point is that “SANs are too slow”. But FVP takes away the pain of managing a server-side flash cache by making it essentially invisible, and you don’t have to forklift upgrade your SAN to get the benefits. The data lives closer to the workload, so the data path from cache to VM has lower latency, removing both the SAN and network I/O hops.

Data-protection is handled by replication to other nodes in the cluster, and there are assertions about this being fine for reliability and performance. If performance is the reason you’re putting this in, then you’ll be on 10Gb networks; 1Gb just won’t cut it, even though they’re supported (again, if they’re on the VMware Hardware Compatibility List).

Sounds like VSAN

This sounds a bit like VSAN, right? Well, sort of, but not really.

VSAN uses a combination of disks and flash installed in the cluster hosts to provide storage; there is no SAN. There are limits to the maximum storage amounts because you can only have 1 flash device per disk group, and a maximum of 5 disk groups per host, and a max of 7 HDDs per disk group. In an 8 cost cluster, that’s a maximum of 1.12 petabytes raw storage using 4TB SATA drives.

VSAN is a smallish flash cache in front of some storage that’s distributed across the cluster, and installed directly in the hosts, which is a lot more like Nutanix than PernixData. Pernix is flash cache in the hosts in front of whatever SAN(s) you already have.

VSAN also doesn’t support all the high-end features you get with a SAN, like advanced snapshotting, replication, fancy VAAI offloads, dedupe, compression, and other fancy stuff. It’s not clear to me if PernixData would get in the way of that, but one would hope not, since it’s basically just a cache.

PernixData Good

If you need flash acceleration for performance or contention reasons, dropping some into your hosts is likely to be easier (and cheaper!) than upgrading your SAN. You might have an older 4Gb FC fabric, and upgrading the fabric would be costly and a pain. If you only need to solve performance for one or two VM clusters, that might not be worthwhile.

Your hosts will need to have the space for some flash, though. It could be SSDs, or PCIe cards, because PernixData will work with both, so that makes it easier to hopefully find a free slot.

You can protect the investment you’ve made in an expensive SAN and FC fabric, and maybe increase its life in a relatively cost-effective way while you figure out what you want to do in the longer term. Or, you can more directly target a specific workload that needs acceleration, but not spend money on something that would be more generally available, but that nothing else needs.

Also, it’s software, so it’s pretty easy to upgrade when new versions come out, and maybe get some performance increases if their cacheing algorithms are improved.

PernixData Less Good

The trade-off here is that you have to put flash in your hosts, and match capacities across the cluster to be able to support appropriate data-replication to keep your data safe. If you only need read-caching, well, most SANs can do that fine, and if yours can’t, you have bigger problems and should re-evaluate your storage architecture.

It also doesn’t help you if you’re using NFS datastores, and you should be looking at something like Infinio instead.

And if you’re using a non-VMware hypervisor like Hyper-V or Xen, PernixData is not for you. PernixData may well decide to support other hypervisors in the future, in the same way that Veeam and others started on VMware and then added Hyper-V later, but it’s not there today, and I haven’t seen any roadmaps saying it’ll be around soon.

Eigenmagic Evaluation

FVP seems like a decent enough product (certainly Chris Wahl seems to like it) and it’s fairly well targeted to a specific problem set. Their focus on a specific problem is good.

Flash cache for SAN backed VM clusters makes a lot of sense, and is a tremendous bridge for those who want some flash to make things faster, but don’t want to rip and replace their entire environment.

There are two things I’d like to see from their roadmap:

  • NFS support. NFS datastores are tremendously useful for a bunch of reasons, and I see no reason a software flash cache can’t support any storage protocol I might choose to use.
  • Support for other hypervisors. Again, probably not too hard if the architecture is right. It grows their potential market, and reduces their reliance on VMware. That’s important now that VMware have VSAN, and we’ve seen they’re willing to throw an elbow with some partners.

I can’t see PernixData surviving as a separate company for 10 years off the back of this one product, though. The niche is good for focus, but they need to keep growing so they can pay off the VCs. I see them getting bought by someone else to round out their product portfolio when the inevitable storage company consolidation happens, rather than going the IPO route.

Bookmark the permalink.

2 Comments

  1. Pingback: SFD5 Prep-work: PernixData

  2. Pingback: PernixData FVP Accelerates All The Things | eigenmagic

Comments are closed