How to Double* the Performance of vSwitch-based Deployments

This was a talk I have at the OpenStack Summit Berlin in November 2018. In this talk, I discussed the NUMA-aware vSwitches feature recently added to OpenStack Nova, explaining what it was, why you’d want it and how you could use it.

vSwitch solutions, such as DPDK-accelerated Open vSwitch, promise to provide much of the performance of SR-IOV with increased flexibility and decreased operational cost. However, while the switch may be virtual, the hardware that the vSwitch and VMs are running on is not. Unlike SR-IOV, NUMA affinity was not accounted for when instances using vSwitch-backed networking. This could result in a loss of up to 50% of network performance.

This changes in Rocky. With the completion of the NUMA-aware vSwitch feature, nova now has the ability to schedule and create you instances in a manner that will deliver the maximum possible performance from your vSwitch-powered network. Come learn how we did it and how you can start using this new functionality in your own deployments.

What can I expect to learn?

  • How NUMA can adversely affect the performance of your vSwitch-based deployment
  • What features nova provides to eliminate this performance penalty and how you can use this functionality in your own deployments
  • The work ongoing to continue improving performance in this area and related areas

More information is available on the OpenStack Summit Berlin website.

comments powered by Disqus