On Thu, Sep 20, 2018 at 08:23:06AM +0000, Daniel McGinnes wrote: > Hi Roman, > > unfortunately Kubernetes seems to be using version 1 cgroups, so I can't > see that stat - I'll investigate if there's a way to get Kube to use V2 so > we can check this.. Hi Daniel! Yeah, it might be not so easy, AFAIK. Alternatively, you can expose this cgroup v2 data in v1 interface using an off-stream patch, just for debugging. Should be pretty straightforward; I can help with it, if necessary. > > There wasn't memory pressure, I just run it in a pretty controlled way > when running the test - so initially it sounds like what I saw was > expected. I then ran stress --vm 16 --vm-bytes 2147483648 which did create > some memory pressure and I saw oom killer getting invoked - it seemed > pretty similar behaviour to before where only a small amount of the "lost" > memory was reclaimed... Maybe I was being too severe with stress and the > memory would be reclaimed at a slower rate under more reasonable memory > pressure? So, did you run the stress -vm after the main workload or in parallel? Can you, please, try to create some ambient memory pressure? Does echo 3 > /proc/sys/vm/drop_caches help to reclaim the memory? Thanks! Roman
