On 2023/03/07 16:49, Fabio Martins wrote: > On Tuesday, March 7, 2023, Claudio Jeker > > > > > > No need to collect flamegraphs, the issue is massive contention on the > > kernel lock because of high IO load. I see similar behaviour with iogen. > > Currently competing read and write calls clash with the async buffer > > handling which also requires the kernel lock to finish their work. So more > > concurrency makes it worse. Fixing this is a major task.
Confirmed; closef, read, doopenat are prominent. https://junkpile.org/kstack-nms2030307.html > Can a ramdisk improve the performance while there are no changes in the > code? Not unless the existing storage is really slow; the issue is kernel contention not disk access speed. MFS doesn't help this.
