| kata-irc-bot | <a.gaevoy231> Hi guys, I have a question, Can I run vms with kata in microk8s like a kubevirt (to use it with a legacy app that I can't move but require keep everything in containers) | 09:23 |
|---|---|---|
| kata-irc-bot | <fidencio> Sorry, I am not that much familiar with microk8s. | 10:03 |
| kata-irc-bot | <fidencio> What does it use behind the scenes? | 10:03 |
| kata-irc-bot | <fidencio> I see that it uses containerd behind the scenes, so it should work, although you may need to configure containerd acconrdingly. | 10:04 |
| kata-irc-bot | <will> Hi guys, anyone had any luck running Kata on Azure EPYC based nodes? We're seeing x5 slower performance vs the Intel counterparts. We think the large number of sperate L3 caches + numa setup could be influencing this. Performance is better than Intel in runc for the specific tests run. Diagram of the CPU/cache setup attached. Any ideas would be appreciated. | 16:06 |
| kata-irc-bot | <feng.wang> Do you use kubernetes for the test? It’s possible to pin the kata vm to a specific numa node through cpu manager/memory manager/topology manager. | 16:24 |
| kata-irc-bot | <feng.wang> Internally we always pin the kvm vcpu threads a fixed set of cpu cores (through our own custom implementation), which gives us better performance. | 16:26 |
| kata-irc-bot | <will> hey Feng, thank you for your responses. We're running it in K3s specifically before running tests on AKS. | 16:51 |
| kata-irc-bot | <will> https://documentation.suse.com/sbp/all/html/SBP-AMD-EPYC-2-SLES15SP1/index.html this caught our eye before | 16:52 |
| kata-irc-bot | <will> Is there any literature about CPU pinning in Kata that you're aware of? | 16:53 |
| kata-irc-bot | <feng.wang> @will if you just want to have a hack solution to verify your hypothesis, you can fork the kata shim code and decorate the qemu command using numactl to make sure the qemu vm is running on the same numa node (assuming your pod can fit within a numa node). Kata runtime itself doesn’t support cpu pinning right now. The challenge is how to allocate CPU cores to different pods on the same node, which requires support from upper layer | 18:42 |
| kata-irc-bot | (e.g. kubelet/containerd). The cpu/memory manager is the closest thing I am aware. | 18:42 |
| kata-irc-bot | <will> thanks, will try set this up over the coming days :+1: | 19:30 |
| kata-irc-bot | <a.gaevoy231> @fidencio can you point me to some guidelines how to run something like legacy vm with kata and kvm? | 21:36 |
Generated by irclog2html.py 2.17.3 by Marius Gedminas - find it at https://mg.pov.lt/irclog2html/!