virtualization.info

Finally, the Standard Performance Evaluation Corporation (SPEC) released the first industry standard benchmark for hardware virtualization and OS virtualization: SPECvirt_sc2010
The standard body is working on this since November 2006.

The benchmark designed to simulate the activity of three typical workloads: a web server, a Java application server and a mail server.
To do so it leverages existing SPEC benchmarks, modified to measure performance in a virtualization environment: SPECweb2005, SPECjAppServer2004 and SPECmail2008.

SPECvirt_sc2010 adopts the same approach used by the VMware benchmark VMmark, the tiles, to measure scalability: the framework deploys additional tiles until overall throughput reaches a peak and all virtual machines continue to meet required quality of service (QoS) criteria.

The web server workload is made by three tiers: a file server, an infrastructure server and of course the web server. They are hosted in two virtual machines (file server and web server are together).
The application server workload is made by two tiers: a backed database server and of course the application server. They are hosted in two VMs.
The mail server workload is made by just a single tier: the IMAP mail server, which is hosted inside a single VM.
A sixth, additional workload called SPECpoll has been created and put into an idle virtual machines. It sends and acknowledges network pings against the idle server in 100% load phase to measure its responsiveness and to all VMs in the 0% load phase (active idle) during power-enabled runs.

The benchmark is also able to measure power consumption and power/performance relationship.

It’s not available for free: it costs $3,000 with discounts for qualified non-profit and educational institutions.

Customers may believe that the availability of this standard finally closes the endless discussions about the value of VMmark, originally introduced more than three years ago, the VMware’s restriction in using it to compare vSphere against other products, and the quality of independent alternatives like the Project Virtual Reality Check (VRC) or the Anandtech vApus Mark.

Unfortunately, it may not be the case yet. The SPEC virtualization subcommittee that developed it in fact includes the following members and contributors: AMD, Dell, Fujitsu Siemens Computers, HP, IBM, Intel, Microsoft, Parallels, Red Hat, Sun Microsystems, Trigence, Unisys, and VMware.
For some reasons anyway, the press announcement issued by the SPEC only includes a subset of the names: AMD, Dell, Fujitsu, HP, IBM, Intel, Oracle, Red Hat, Unisys and VMware.

Microsoft is not there. And Citrix is not even a member of the subcommittee. It’s unclear if the two support this benchmark or not.
If they don’t, then that customers will get only some benefits: now will be able to use SPECvirt_sc2010 to compare multiple hypervisors, but until all players will validate the benchmark, the results will always be questionable.

Update: IBM is the first and only OEM that already published some SPECvirt_sc2010 scores, and they are quite interesting because the benchmark has been used to measure the performance of KVM.

The System x x3650 M3 achieved an overall performance score of 1169 @ 72 VMs served by Red Hat Enterprise Linux (RHEL) 5.5 and KVM.

The x3650 M3 was configured with the Intel Xeon Processor X5680 (3.33GHz with 256KB L2 cache per
core and 12MB L3 cache per processor—2 chips/12 cores/6 cores per chip), 144GB of memory and ninety-six 73GB disk drives.