On Tue, 13 Apr 2021 06:14:42 GMT, David Holmes <dhol...@openjdk.org> wrote:
> This strikes me as the wrong fix to the problem. isCpuSetSameAsHostCpuSet is > only intended to be used as a simple optimization when the configured cpuset > happens to match the hosts. What you are looking for is a fix to the problem > when there is no cpuset set at all. I pushed new commit to move the fix into `getCpuLoad()`. > It strikes me that in getCpuLoad() if there are no quotas and no > effective-cpu-set and no cpusets.cpus value, then it should fallback to using > the host values rather than returning -1. That said, the problem may also be > that we have a containerMetrics object when no container is actually active! > Perhaps that is the true bug here? I think this problem is similar to [JDK-8264482](https://bugs.openjdk.java.net/browse/JDK-8264482) (PR #3280 ). Currently both HotSpot and MBean check whether cgroups exists, but the the host (not into the container) might have it. We've discussed about it in #3280 , then I think it is difficult to modify to check whether the VM is in the container. Thus I think it is reasonable to just return host CPU load value at here. ------------- PR: https://git.openjdk.java.net/jdk/pull/3447
