Open Source Summit + Embedded Linux Conference Europe 2020

Being able to craft a detailed virtual topology for a VM may be crucial for achieving good performance. But it is also risky, as interfaces become more complex, and an inconsistent configuration may be selected, causing more harm than good.

E.g., it would be good to be able to specify the size of caches, for cases when some software (e.g., glibc) inside the VM checks it and decides whether or not to enable some optimizations depending right on that.

On the other hand, even just defining the vCPUs topology (threads, cores, NUMA nodes, etc) may lead to less stable or outright worse performance, if the vCPUs and the memory of the VM are not properly pinned at the host level.

In this talk, we will show some first-hand examples, we will outline what is currently there in Linux, libvirt and QEMU and we will discuss if it is possible to improve things even further.