ceph-iosched

ceph-iosched

#!/bin/sh

# prefer cfq (el7), then mq-deadline (el8), then do nothing

if grep -q cfq /sys/block/sd*/queue/scheduler; then
  # tune SSDs to use noop scheduler and spinning disks to use cfq
  for DISK in /sys/block/sd*; do grep -q 0 ${DISK}/queue/rotational && echo noop > ${DISK}/queue/scheduler; done
  for DISK in /sys/block/sd*; do grep -q 1 ${DISK}/queue/rotational && echo cfq > ${DISK}/queue/scheduler; done

  # tune cfq not to penalize writes when reading heavily
  find /sys/block/sd*/queue/iosched/group_idle -exec sh -c 'echo 0 > {}' \;
  find /sys/block/sd*/queue/iosched/slice_idle -exec sh -c 'echo 0 > {}' \;
  find /sys/block/sd*/queue/iosched/quantum -exec sh -c 'echo 32 > {}' \;
  exit 0
elif grep -q mq-deadline /sys/block/sd*/queue/scheduler; then
  # tune SSDs to use none scheduler and spinning disks to use mq-deadline
  for DISK in /sys/block/sd*; do grep -q 0 ${DISK}/queue/rotational && echo none > ${DISK}/queue/scheduler; done
  for DISK in /sys/block/sd*; do grep -q 1 ${DISK}/queue/rotational && echo mq-deadline > ${DISK}/queue/scheduler; done

  # tune mq-deadline not to penalize writes when reading heavily
  find /sys/block/sd*/queue/iosched/write_expire -exec sh -c 'echo 500 > {}' \;
  find /sys/block/sd*/queue/iosched/writes_starved -exec sh -c 'echo 1 > {}' \;
  find /sys/block/sd*/queue/iosched/fifo_batch -exec sh -c 'echo 8 > {}' \;
  exit 0
else
  echo Cannot find supported io scheduler.
fi

Still using bfq on all el8 osd servers?:)

Still using bfq on all el8 osd servers?:)

yes we are. do you have a better suggestion?

Nope, we chose as well bfq. My question for double check for right choice. Thanks.

Seems cfq has been gone in kernel 5. (https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=0e9da3fbf7d81f0f913b491c8de1ba7883d4f217)
Will you use bfq or mq-deadline instead?

bfq

Updated -- in 2022 we switched to mq-deadline.