We discovered that on our systems with lustre mounted, the kworker is using signifficant amount of CPU.
perf top shows on an idle system:
We currently have grant_shrink=0 on this system.
Looks like doing just du -hs /fs makes the problem go away for some time.
Also unmounting the filesystem makes the problem go away.
This is Centos 7.6 system with Lustre 2.12.0