After a bit of a hiatus in properly maintaining this site due to various maddening deadlines and the stress that comes with finding financing for the coming few years (we succeeded, by the way!), starting today, this site will be receiving weekly updates, so stay tuned!
As for today's update, it's time to share a bit of info on the things we learned at this year's VMworld Europe in Cannes. As we 4-manned the event, we were able to attend quite a large amount of sessions. A full-sized article on that is coming up on Anandtech, but for our own website, we'll put up a couple of our notes to start out with:
Performance Best Practices (a session by Scott Drummonds)
We didn't learn quite as much as we'd hoped from this session, but its contents are nonetheless important considerations for anyone hoping to virtualize their data center. We'll list the most interesting pointers down here:
- Any application that spends most of its time using the CPU and not much of anything else will barely feel the impact of virtualization. Performance is impacted primarily by applications that require a lot of time spent switching to kernelmode.
- The highest measured network speed using ESX is 16Gb/s
- ESX is able to handle at most about 100k IOPS and 600 disks in total.
- Today's CPU's are much better equipped to handle virtualization than those of a few years ago. Consider upgrading for a very large performance boost.
- The use of jumboframes is recommended for every NIC. Separate NICs should always be used for interserver communications within the data center, like VMotion.
- Disk caching and properly distributed disks are an important factor for storage performance. The newest version of ESX will improve a lot on iSCSI performance.
- VMware recommends always using VMFS with 64k boundaries. This is the standard setting in vCenter. If you are not using vCenter, remember to set it yourself.
- Use an OS that causes less interrupts, and disable every device you don't need.
- Enable large pages for the TLB, to reduce cache misses.
- Applications that do not scale well on multicore platforms natively might do well in a multiple VM-deployment. This way, each app can perform at its most optimal setup.
If you're interested in learning more, stay tuned! We'll be updating this blog regularly with interesting tidbits of information.