Rich McDougall and Scott Drummand, performance gurus over at VMware, presented a fantastic review of some of the performance testing VMware has been doing at the VMware analyst day. I posted something about the findings of that research in the post, Is VMware busting the virtualization tax myth. I've just gotten an impressive update. They have done some elaborate testing and can dispell many of the most common myths including the following:
- Applications always run slower in a virtual environment
- Applications having high I/O aren't good virtualization candidates
- Virtual machine software have bottlenecks preventing multiple copies of large applications concurrently.
- VMware doesn't leverage hardware assists
- VMotion, VMware's Distributed Resource Scheduler (DRS) and memory sharing aren't really useful in large scale deployments.
Although I don't have the space to present all of the data, analysis and colorful figures here, I was impressed. Having executed benchmarks and performance studies while I was with Digital Equipment (may they rest in pieces), I found their approach both reasonable and, in all likelihood, supportable.
I also enjoyed their commentary on why some organizations don't see similar performance in their environment. Most of the issues come down to improperly configured VMs (insufficienct number of cores/processors allocated, memory or storage), improperly configured physical systems (the usual suspects: processor power, memory or storage) or insufficient network bandwidth. Network-based applications usually experience bottlenecks in the following order:
- Network bandwidth and latency
- Storage performance, latency and capacity
- Memory capacity
- Processor capacity or performance
VMware offers a number of interesting studies on performance on their website. If you're interested in this subject, I'm sure you'd get some valuable information by reading those papers.