Information and announcements from Program Managers, Product Managers, Developers and Testers in the Microsoft Virtualization team.
Hi, I’m Alex Miroshnichenko, chief technology officer of Virsto Software. One of the most exciting events of my summer was the announcement of Microsoft Windows Server 2008 R2 Service Pack 1 (SP1). Though still in beta, SP1’s stability has lived up to production standards in Virsto’s test labs.
SP1 brings Dynamic Memory to Microsoft Hyper-V. In practical terms, you should be able to run meaningfully more virtual machines on the same server hardware, since static memory limits are gone.
Dynamic Memory also eliminates a key argument that competitors have held against Hyper-V. We should not expect the competition to stop the FUD completely. However, it will become increasingly difficult to use the “memory overcommit” argument against Hyper-V being a true enterprise grade virtualization solution.
Why would a storage guy like me be so excited about dynamic memory? It is rather straightforward: now that memory is no longer an issue for increasing virtual machine (VM) density per physical host, the next major obstacle becomes the storage architecture. The technology we have developed at Virsto Software removes that obstacle.
Virtualized hosts drive storage traffic differently than physical servers because they run multiple instances of independent operating system (OS) images. We call this effect the “VM I/O Blender”. Each OS instance optimizes its I/O patterns on the assumption it owns the hardware; however, the virtual machine hardware itself is virtual. The I/O streams are “blended” in the virtualization layer, making the I/O pattern through the physical storage interconnect highly random. As we all know, storage devices are much worse at random I/O than sequential – as much as two orders of magnitude slower.
The more virtual machines per physical server, the more random the storage I/O stream becomes, and the more the VM I/O Blender will hinder server performance. Half, even 80%, of a server’s storage throughput can be lost.
How might you solve this problem? The simplest (and may I say the worst) way is to throw money at the problem: buy expensive storage. And when I say “expensive” I mean it. A smarter way is to install a storage software solution that is specifically designed to handle the unique storage I/O patterns and lifecycle of virtual machines. By removing storage bottlenecks, Virsto One software for Hyper-V can deliver three to four times the effective VM density while providing other features essential for VM storage management.
The VM I/O Blender is an artifact of all hypervisors, easily demonstrated on any virtualization platform. We’re pleased to be working with Microsoft to obliterate it on Hyper-V. We have performance comparisons using Hyper-V with Virsto One versus competitive hypervisors that spew about “memory overcommit advantages”: Hyper-V is several times faster. Try it for yourself.
Nice, you bring what high-end and pricy storage systems do for years to low-end and cheap storage systems or DAS don't do at all. This plug-in is great idea indeed. Does it work with VMware?
"Hyper-V is several times faster. Try it for yourself."
Where could one obtain details about the environment and workload used to perform this comparison between the "competitive hypervisor" and Hyper-V+Virsto One?
"The first rule of VMware benchmarks is that you don't talk about VMware benchmarks. The second rule of VMware benchmarks is that YOU DON'T TALK ABOUT VMWARE BENCHMARKS."
@PiroNet: I'll ask Alex from Virsto to reply to you.
@Eric Gray: I'll ask Alex from Virsto to reply to you, too.
@OhGeez: let's keep the sarcasm polite. A smiley face helps ;-)
Thanks PiroNet. Virsto One runs only on Hyper-V today. Based on customer demand, we will deliver Virsto One on other hypervisors. We haven’t said which ones nor have we said when. The good news (for Virsto) is that there are plenty of storage I/O problems on all the virtualization platforms!
Eric, my claim was that Hyper-V with Virsto One deals with the VM I/O Blender better than other hypervisors, and that in certain benchmarks is several times faster. The primary point of my post was that when hypervisor vendors alleviate memory bottlenecks, storage bottlenecks become much more visible, which is where Virsto One shines.
You are of course aware of the restrictions certain vendors put on publishing performance benchmarks. For this reason, I don’t believe we are allowed to publish an answer to your question. If I’m mistaken, please contact me directly to educate me.
I would be happy to show any virtualization user how they can reproduce the results themselves.
Thank you for the response. In order to remove benchmarking EULA issues from this discussion, it should be feasible to simply describe the test environment and the workload. This seems useful to anyone - pure Hyper-V shops included - considering Virsto One.
I am looking forward to it, and appreciate your offer to show how anyone can reproduce the results.
Thanks for the followup. Rather than post a lengthy response here, I invite you to please follow me over here: