Sorry for the delay, but here’s some extra details.
The issue is that processes are reporting might higher CPU than they’re actually using. Here’s output from graphite. app-7 and app-8 are CentOS 7.4 64bit containers running on a lxd 3.11 host, with lxcfs 3.0.3 pulled from master about 1 month ago.
Even when there’s ZERO load on the boxes, processes are reporting that they’re using more CPU then they actually are.
Here’s the output from top on app-7: https://pastebin.com/Wh5x85JQ - Even though the summary line is reporting 100% idle, java is reporting 200% used and top 100% used. Even though there’s no actual load on the box.
Here’s the output from top on a container spun up from the lxd images repo: https://pastebin.com/Whh6T7Ha - same issue, with a vanilla centos 7 container spun up from images:centos/7/amd64 - top process reporting 100% cpu, even though it’s not using it.
Here’s the output from htop on app-7: https://pastebin.com/JPhCv9dn - It does the same thing, reporting cores are fully loaded when the process is actually idle, or only using a few % of CPU, not hundreds of %.
I’ve tried building lxcfs from master today, same issue. I’ve tried building 2.0.8, and while I can trigger the same issue, it seems to happen less frequently.
I’m not sure how to debug this further. But so far, the issue only exists on CentOS 7 containers. I cannot reproduce it on CentOS 5, 6 or Debian containers.
Any help would be appreciated. Thanks in advance!