Hyper-V RC1 Released as Microsoft Shares Performance Data
Published: May 28, 2008
by Alex Woodie
Microsoft last week delivered a second release candidate for Hyper-V, the new hypervisor that will enable Windows Server users to carve their boxes into numerous virtual machines. Hyper-V Release Candidate 1 (RC1) introduces some new features, such as support for Windows 2000 and an easier installation process. But it's Microsoft's experience in using Hyper-V to run some of its Web sites, and the early glimpse it provides into Hyper-V's performance characteristics, that is most interesting.
By all accounts, Microsoft is still on track to deliver Hyper-V, as it promised, 180 days following the release to manufacturing (RTM) of Windows Server 2008, which would place Hyper-V's debut sometime in August, if not sooner. But it's still not done tweaking the software, which is critical to Microsoft's success in the virtualization market.
Like the first release candidate, which Microsoft is now calling RC0 (previously referred to as RC), RC1 is also "feature complete," according to Microsoft. Obviously, RC0 was not feature complete, or Microsoft wouldn't have added new features with RC1. In any event, it's good to see Microsoft respond to the testers' feedback and add capabilities based on their input. But we just can't take Microsoft at its word that no other features will be added for a third release candidate, or RC2.
So what's new with RC1? Probably the most significant feature is support for Windows 2000 as a guest operating system running under Hyper-V. The software will support Windows 2000 Server SP4 and Windows 2000 Advanced Server SP4 operating systems as single-processor servers. This is significant for anybody still running old Windows 2000 applications and who would like to consolidate these environments onto newer hardware.
The second new feature is a simplified installation process for running Windows Server 2008 as guest environments on Hyper-V. Microsoft worked on making the process of installing the integration components "simple and consistent." The final new feature is support for mice when running Novell SUSE 10 as a guest operating system under Hyper-V. This piece has been added to the integration components, Microsoft says. Other operating systems certified to run as guests under Hyper-V, you will remember, include Windows Server 2003 SP2, SUSE Linux Enterprise Server 10 SP1 (X86 and X64 versions), Windows Vista SP1, and Windows XP SP3.
The current beta of System Center Virtual Machine Manager 2008 (SCVMM) is not compatible with Hyper-V RC1, according to Microsoft, which recommends that customers who are testing the SCVMM 2008 beta should continue evaluating it with Hyper-V RC0. The lack of compatibility raises concern for these products, which are heavily dependent on each other (SCVMM provides the human-facing interface for managing Hyper-V), but Microsoft undoubtedly will fix the incompatibility problem before either product RTMs.
Microsoft also released the first glimpse into Hyper-V's performance in a report that's available from the Microsoft Web site as a Word doc here. The report describes the experience that Microsoft's operations team has had using Hyper-V to run the Web and database servers powering its TechNet and MSDN Web sites.
According to Microsoft's Rob Emanuel, the software has worked as advertised. "Possibly the most surprising finding was that Hyper-V was far more stable than we had expected for a beta version deployment," Emanuel writes in a blog posting here. "There was in fact no difference we found in stability or availability between Hyper-V and a physical deployment. We were also not able to identify any bugs for the Hyper-V team during our deployment, under either full production load or even stress load."
Microsoft measured the performance overhead levied by Hyper-V during the MSDN and TechNet Web server test. According to Microsoft's workload and performance analysis, virtual machines running under Hyper-V consumed 5 to 6 percent more CPU than the parent partition, "with linear progression as the number of requests increased," Microsoft noted in the report.
In terms of total throughput, the virtualized Hyper-V system started to lag the performance of a native physical system as Web site traffic increased on the MSDN site. Compared to running the Web site on a Hyper-V-managed virtual machine environment, a physical environment was able to process 21 percent more "requests per second per 1 percent CPU," which is a virtualization performance metric we're going to hear more about.
According to Microsoft's data, when the "relative server load" for the Web site hit 80 percent, the virtualized system was running at about 66 percent CPU utilization, while the equivalent work done on a physical environment was only consuming about 50 percent of its CPU resources. In any event, Microsoft says the data points to about a 3 percent performance hit when applications running under Hyper-V are "oversubscribed," or pushed to their limits.
As Microsoft works with Hyper-V, it will be playing around with performance figures and trying to develop a model that provides a good reflection of the types of performance characteristics its customers are likely to encounter as they roll out Hyper-V in their environments (and also a model that reflects well upon Hyper-V, undoubtedly). "The model will likely include the current requests per second per 1 percent CPU metric with additional memory and I/O performance characteristics as we virtualize other Web applications and gather more test and production data," the company says.
One element of Microsoft's Hyper-V setup that didn't help performance was its decision to deploy the Web site virtual machines as dynamic virtual hard disks (VHDs). Microsoft says dynamic disks don't perform as well as fixed or physical disk options, but they do provide the flexibility to move them. Microsoft plans to test all disk options in the future, when it implements a storage area network (SAN) with Hyper-V.
As more people start using Hyper-V, it will be interesting to see how its performance compares to other hypervisors, notably VMware's ESX Server. Last year, VMware launched a new benchmark, called VMmark, for measuring hypervisor performance on various workloads. Standard Performance Evaluation Corporation, or SPEC, is also developing a virtualization benchmark; the work is still in subcommittee.
Microsoft Gets 'Feature Complete' Hyper-V Out the Door
'Viridian' Hypervisor Gains Formal Name: Hyper-V
'Viridian' Hypercall APIs to be Open Source, Microsoft Says
VMware Puts Virtualization to the Test with VMmark Benchmark
Microsoft Slashes 'Viridian' Features to Meet Ship Date