Comeback time for the storage hypervisor

When two separate companies, Virsto and Datacore, announced the release of 'storage hypervisors' almost simultaneously, and VMware joined in with the release of its vSphere Storage Appliance (VSA) software, it was clear that something was up.

For long regarded by many as an answer in search of a question, storage virtualisation is now targeting the cloud, and is being touted as a key performance booster for VDI and other cloud-related infrastructure.

Why is it so important to the cloud? “People have been commenting that VMworld is turning into a storage show – storage really is the impediment for everything about cloud,” says former VMware staffer Gregg Holzrichter, who is now VP of marketing at storage virtualisation start-up Virsto.

He argues in particular that virtual machines (VMs) have changed the pattern of disk I/O within servers, adding that once you have 10 virtual servers – or for VDI, 100 desktop images – all on the same physical machine, their collated IO becomes too mixed and random, leaving the host's file system struggling to keep up.

"Access tends to be more random now because there are more applications on the host," confirms Carlos Carreras, VP of alliances and business development at storage virtualisation pioneer DataCore. "From our perspective, it is the same I/O pattern but multiplied by a factor to represent multiple servers." 

He adds: "With VMs it might be time to look at your back-end and maybe bring in new technologies such as flash and SSD. Storage virtualisation software makes it easier to do that – it adds flexibility, and VMs heighten the need for flexibility in the back-end."

The idea of virtualising and pooling storage has been around for a very long time – technically, the way an IDE hard drive allocates disk blocks is virtualisation, as is RAID – but as a separate software layer it dates back perhaps a couple of decades. It has been implemented in different places - as a server plug-in, as an appliance dropped into the storage network, or within a disk array such as an EMC Symmetrix, IBM Shark or Hitachi USP - but they all do the same key thing, which is to break the tie between the logical disk volume and the physical disk drive.

If that sounds familiar, it ought to, and while calling storage virtualisation a hypervisor is undoubtedly a marketing gimmick, it is also a fair description, says analyst Jon Toigo, of Toigo Partners International. He points out that what virtualisation does for storage is a close analogue of what a hypervisor does for server resources: it adds a layer of abstraction, decoupling the logical from the physical.

“It is a hypervisor but for storage applications, not business applications. It is a rebranding, but it points more to a software layer that runs as a package, allocating and reallocating resources,” he says.

So just as VMware, Hyper-V and Xen abstract the server into a set of resources which can be allocated to various VMs, software from the likes of DataCore and Virsto turns your storage resources into a networked pool of blocks which can be recombined in all sorts of ways. Several drives could be combined into one giant volume, say, or the same data written to two mirrored locations, or a logical drive can be thin-provisioned meaning that the operating system sees its full capacity, but it only actually occupies as much physical disk space as it contains data.

Virsto's solution uses storage virtualisation at the hypervisor level. The VMs see a single virtual disk volume, but in fact they initially write to a tier of solid-state disk (SSD) acting as a sort of cache, Virsto's software then optimises the data before writing it out to spinning disk.

“We are hypervisor-based, and the central thing for us is the VM, not the array or LUN,” explains Holzrichter. “I think there is a fundamental world-view difference between storage-centric and server-centric. We were conceived after VMs and saw the challenge of very random IO from server consolidation.”

He adds: “VDI is a specific use case where all the factors are exacerbated – it's 100 desktop images, not 10 VMs, so the storage IO is even more random. Many VDI projects fail because the IO isn't there, despite the preliminary IO calculations.”