You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The very newest Kubelets (v1.19.0-alpha.1 - v1.19.0-beta.1, latest at time of writing) have trouble couting vCPUs on some OSes (incl. Fedora CoreOS), related to updates to cadvisor with CPU topology features. Within those versions, vCPU is detected as 0 (definitely wrong).
A candidate fix detects the number of cores, unaware of disabled SMT, which I suspect is still wrong. I wanted to bubble this ongoing discussion to Fedora CoreOS folks in case someone has ideas on correctness.
I think that root-cause was a fact that cAdvisor has always been ignoring online/offline CPU information. It must be taken into account now and I believe that it should solve the problem. See: google/cadvisor#2567 (comment)
Considering the discussions going on in the upstream issues can someone summarize and let us know if there anything for FCOS to do in the context of this bug?
Great info @iwankgb thanks. AFAIK, this has been narrowed down to cadvisor changes that detect effective vCPU count as 0, which seems to stem in whole or part from not handling OSes that disable simultaenous multi-threading (SMT) (Fedora CoreOS, Google's internal distro).
cadvisor currently reports the wrong CPU count on FCOS, they're discussing fixes, and that would eventually get merged and then included in a Kubernetes v1.19 beta or rc. I don't think FCOS would need to make any changes, I can close this out.
The very newest Kubelets (v1.19.0-alpha.1 - v1.19.0-beta.1, latest at time of writing) have trouble couting vCPUs on some OSes (incl. Fedora CoreOS), related to updates to cadvisor with CPU topology features. Within those versions, vCPU is detected as 0 (definitely wrong).
A candidate fix detects the number of cores, unaware of disabled SMT, which I suspect is still wrong. I wanted to bubble this ongoing discussion to Fedora CoreOS folks in case someone has ideas on correctness.
The text was updated successfully, but these errors were encountered: