On Wed, Sep 08, 2021 at 08:15:11PM +0800, Yi Tao wrote: > In a scenario where containers are started with high concurrency, in > order to control the use of system resources by the container, it is > necessary to create a corresponding cgroup for each container and > attach the process. The kernel uses the cgroup_mutex global lock to > protect the consistency of the data, which results in a higher > long-tail delay for cgroup-related operations during concurrent startup. > For example, long-tail delay of creating cgroup under each subsystems > is 900ms when starting 400 containers, which becomes bottleneck of > performance. The delay is mainly composed of two parts, namely the > time of the critical section protected by cgroup_mutex and the > scheduling time of sleep. The scheduling time will increase with > the increase of the cpu overhead. Perhaps you shouldn't be creating that many containers all at once? What normal workload requires this? thanks, greg k-h
