Virsto One, Hyper-V and the I/O Blender effect
One of the things I’ve come to love about blogging is the fact that I occasionally get contacted by the odd tech start-up. Keen to demonstrate their latest market leading idea that is going to revolutionise the industry as we know it. Earlier this month I was contacted by Mindy Anderson who is the Product director at tech start-up Virsto (short for Virtual Storage). Virsto had a new product for Microsoft Hyper-V that they wanted to demonstrate to me in advance of their big product launch. Having looked at Mindy’s background in the storage industry I was very keen to hear more about their new product.
The product is called Virsto One and is aimed solely at Windows 2008 R2 Hyper-V. The product introduces some new features like thin provisioned clones & snapshots, that expand the functionality of the standard Hyper-V product. The most interesting feature in my opinion though is the attempt to tackle the virtualisation / storage problem commonly known as the I/O blender effect.
So what does Virsto One look like?
The software itself installs in the parent partition of each Hyper-V host and consists of the filter driver, a system service and a VSS provider. The filter driver sits above the raw storage (any block storage) and presents a VHD object to the parent partition. This setup allows users to configure physical storage once and then use Virsto One to carry out all future provisioning tasks. This includes full support for creating thin provisioned, high performing, cluster aware snapshots and clones from either the Virsto One Hyper-V MMC snap-in or Powershell.
So what about the I/O blender effect?
Most storage technologies are not designed for the virtual data centre, most are still designed around the one to one physical server to storage model. Think of a number of virtual machines all with predictable I/O behaviour (if you think of them as physical). What tends to come out of the physical hypervisor host is a large amount of completely random I/O. Random I/O has an obvious performance impact when compared with sequential I/O so as you increase the number of VM’s you increase the random I/O from your Hyper-V host. So as VM density increases performance drops, as we all know low VM density is not your objective when you embark on a virtualisation project.
So Virsto One has an interesting way of dealing with this. Although the “secret sauce” has never been divulged in-depth in its basic form they journal the random I/O that comes down from the Hyper-V host to staging disk. A staging area is required per physical Hyper-V host and about 20GB / 30GB of disk should support multi-terabyte write downs through use of de-dupe technology. Periodically the data in the staging disks will be flushed / written down to the primary storage location, at this point the Random I/O is laid down sequentially on primary storage to improve read performance. Virsto indicated that in time they would look to support multiple de-stages so that data could be de-staged to another array for business continuity purposes or to the cloud for disaster recovery purposes.
Are there any performance figures to back this up?
Performance figures from the Virsto test lab show the I/O Blender effect in full effect as VM density increases in the standard Microsoft setup. With the Virsto software sitting in the middle, staging the data and de-staging it sequentially, there is an obvious improvement in performance. These test results were from Virsto’s own lab and I stressed the importance of having these independently benchmarked by customers or an external consultancy. Wendy indicated to me that this was something they were looking into, I look forward to reading and sharing the whitepaper when it is eventually produced.
So who would be interested in a product like this?
Well ideally the product would benefit Hyper-V customers who require high density, high performing virtual environments. Hosting companies making use of Hyper-V for selling virtual server instances may well see Virsto as a good way of increasing performance and reducing costs through the use of golden images, snapshots, etc. Who knows though, individual companies with an investment in Hyper-V may well see the benefit in this kind of product. In a way I see it’s not to dissimilar to a company buying PowerPath/VE to increase I/O performance in a vSphere environment.
It is important to note that although this product has been initially built for Microsoft Hyper-V the principals behind it are hypervisor agnostic. I asked the question “why Hyper-V?” at the start of my chat with Virsto, the answer was that Hyper-V had functionality gaps and was easier to integrate into. VMware on the other hand is a more mature product where VMFS has gone some way to deal with the traditional virtualisation storage problems. Citrix virtualisation customers will be happy to hear that testing has already begun in the lab with a version of Virsto one for XenServer, ETA unknown at this stage.
So how much does all this cost?
At the time of the interview, which was a good few weeks back the per socket price being talked about was $1,000 - $2,000 USD per socket, again not to dissimilar to the pricing for EMC PowerPath/VE.
Conclusion?
My impression at the time of the demo and interview was that this was an interesting product and very clever idea. The main selling point for me was the increase in performance, if it can be independently verified you would think the product will simply sell itself. I look forward to hearing more about Virsto in the future and I am particularly interested to see what they can do for other hypervisors especially VMware vSphere with it’s new storage API’s.
