If you thought measuring and maximising disk performance for a single server was complex, wait until you start to research the topic for VM host servers.
This thread was was prompted by a conversation yesterday with Virsto, a company that makes a plug-in for virtualisation hosts with the aim of making your disk systems run faster. The company claims 3-5x -- sometimes even 10x improvements -- in disk performance. Yeah right.
With $12 million from venture capitalists swelling its coffers, Virsto has launched a marketing offensive to back its software technology which, although it runs only on Microsoft Hyper-V right now, is promised for mass-market VMware's ESX in the second half of 2011 -- probably at VMworld which starts in late August.
So why develop for Hyper-V before market leader VMware? According to Eric Burgener, VP of product management, Virsto found that Microsoft was "more willing to cooperate with small company". The aim is "to get out on the main hypervisors including Citrix", and the company "will consider" supporting others.
The problem Virsto is trying to solve is poor disk performance resulting from the way that a server hosting multiple VMs writes highly random data to disk. As we know, disks -- even SSDs -- are not good at random access but run much faster when writing data to sequential areas of the medium. That's what Virsto's VSI product is aiming to do: sequentialise (ugh) random writes and so boost performance. It costs $2,800 per host and around four to six percent additional CPU load.
It works by grabbing the VM host's disk writes -- it looks to the host like a virtual disk -- and writing the random data to a log on the disk. This is then sorted into sequential order and rewritten to disk, along with the appropriate metadata. It's also said to support VM mobility so that as a VM moves from host to host, the storage log follows it.
It also helps to cut storage costs by reducing the need to buy more spindles in order to get performance, which often results in lots of spare capacity, says Virsto, adding that it liberates storage admin time too as they don't have to spend more time provisioning LUNs, as it's easy for the server admins to do it.
I haven't tested Virsto VSI but I have looked around and can't find anything quite like it -- and while I'm sure there are drawbacks, without full-on testing it's hard for me to say. All I can say is that the idea looks promising and may be worth checking out if disk performance of your VM hosts is an issue.