vSphere Tops Hyper-V in Disk I/O

As a user of server virtualization tools, and a follower of the market, I've always thought that VMware's lead was evident primarily on the management side of the ledger--with vCenter integration, vMotion, Distributed Resource Scheduler (DRS) and the like. Microsoft Hyper-V, I thought, was a competitive hypervisor that needed a bit more age in the bottle to develop the features that I love in vSphere. Then Overland Storage hired my firm, DeepStorage.net, to test its SnapSAN S2000 in both environments, and I saw that the same hardware could perform about 23% more input/output operations per second (IOP) under vSphere than under Hyper-V.

The same S2000, connected to the same Dell R710 server, could handle 733 aggregate IOPs under Windows Server 2008 R2 with Hyper-V and 955 aggregate IOPs under vSphere when running our multiple workload benchmark.

I was frankly somewhat surprised at the results, but since both tests ran on the same system, and used the operating system/hypervisor's native iSCSI initiator and multipath support, I don't have a better explanation than that vSphere has a cleaner I/O path than Windows. Given that the performance difference was greatest for JetStress, I'm thinking that I/O latency is the biggest factor in the result.

I'm sure some Microsoftie, or MS fanboi, will respond to this post with tuning tips for Hyper-V, and I'll be grateful for them. However, the simple truth will remain that, in DeepStorage.net tests of high I/O workloads, vSphere supported more virtual machines per host than Hyper-V.

The multiple workload benchmark is designed to expose how storage systems will perform in real-world virtual server hosting environments. It challenges the storage system to not just deliver a reasonable number of IOPs, but to actually support multiple requests from different virtual servers at the same time.  Like VMware's VMark benchmark, our multiple workload test is based on tiles. A single tile consists of:

The object of the test is to determine the maximum level of load the test system can handle while still delivering data within JetStress' 20ms maximum latency limit. The load presented by each system is adjusted until we reach the highest level at which JetStress still passes the test.

The load presented by the IOmeter systems is controlled by adjusting the number of I/Os each has outstanding at any given time, while JetStress is controlled by the number of simulated heavy (.45 IOPS/user) users. 

This data is the result of research and testing for which Overland Storage paid DeepStorage.net. The full report with more detailed test data is downloadable from Overland Storage's Website.