Open andreeaflorescu opened 4 years ago
The benchmarks are pretty useful here just to get some head start into what to expect. We can try similar benchmarks in high density environments (1000+ uVMs) where the main focus should be CPU oversubscription.
FYI: stress-ng has cpu-cache
class, which might be available to measure the performance impact.
https://manpages.ubuntu.com/manpages/xenial/man1/stress-ng.1.html
Our setup of cache topology (at least on x86_64) is kinda dummy.
At the end of the Automated Performance TestingFor Virtualization with MMTests there are some slides with performance benchmarks where they show the difference between host, VM with no topology, VM with topology and VM with a tuned topology.
We should measure this in Firecracker as well and see if we should also tweak our cache topology.