Virtualised environments allow you to run multiple virtual machines on a single server, but how much of a performance hit does each of these individual environments take compared to running on dedicated hardware? A new research study suggests the gap is relatively small and more than compensated for by overall performance gains.
Image courtesy of Shutterstock
The study, conducted by ESG Labs and commissioned by Microsoft, was designed to examine the performance of Hyper-V on Windows Server 2012. Hyper-V has been significantly upgraded in Windows Server 2012, but as it's a relatively new environment, we haven't seen a lot of real-world examples of how its performance works at scale. 22 per cent of managers surveyed by ESG identified performance as a concern when deploying VMs.
The first figure that jumps out? In one test, ESG Labs compared the performance of an OLTP workload running on a Hyper-V virtual machine with one running on dedicated hardware. As you'd expect, the dedicated machine offered improved performance, since it doesn't require any overhead to run the hypervisor. However, the difference in performance was only 6 per cent (1325 transactions per second on the dedicated machine, 1245 on the virtual machine). While there are undoubtedly environments where that 6 per cent might matter, plenty of organisations would be more than happy with 1245 transactions per second.
A similar pattern can be seen in overall response times:
That said, the whole point of virtualisation is to be able to run multiple instances, so we also want to know what happens when additional VMs are deployed. It's well-known that performance improvements don't scale exponentially as you add more virtual CPUs, since the hypervisor workload becomes higher. However, ESG's testing suggests that the efficiency gains when deploying additional VMs are well-worth considering, provided your base hardware is powerful enough. When measuring SQL batch request performance, a 4-CPU configuration handles roughly 250 requests a second. The 64-CPU version handles just under 3000 — a 12-fold improvement.
You would hardly expect Microsoft to release a report suggesting virtualisation sucked. Nonetheless, this test data suggests that arguments that virtualisation leads to unacceptable performance are going to become less common.