Thank you very much, chemal. I confirm your test results. With CFQ, oowriter now doesn't wait until 'dd' is finished. Of course, it is slower to start than normal, but it starts. And the copying doesn't completely hog the system now.chemal wrote: ↑2018/09/24 01:37:37Adding CFQ on top seems to solve both problems. I no longer have to wait until dd is finished, oowriter starts in parallel. The remedy in later kernels is called writeback throttling (4.10+) but RH did not backport this. However, CFQ does something similar: https://patchwork.kernel.org/patch/9661671/Code: Select all
# don't buffer too many writes echo $((16*1024*1024)) > /proc/sys/vm/dirty_background_bytes echo $((48*1024*1024)) > /proc/sys/vm/dirty_bytes # enable completely fair queuing echo cfq > /sys/block/sda/queue/scheduler
Did RH backport that kyber scheduler or is there any other way to try it with the current distro kernel?