Hi, Paul!
在 2022/08/31 19:59, Paul Menzel 写道:
Dear Yu,
Am 30.08.22 um 03:09 schrieb Yu Kuai:
在 2022/08/29 21:58, Paul Menzel 写道:
Am 29.08.22 um 15:14 schrieb Yu Kuai:
From: Yu Kuai <yukuai3@xxxxxxxxxx>
patch 1 is a small problem found by code review.
patch 2 avoid holding resync_lock in fast path.
patch 3 avoid holding lock in wake_up() in fast path.
Test environment:
Architecture: aarch64
Cpu: Huawei KUNPENG 920, there are four numa nodes
Raid10 initialize:
mdadm --create /dev/md0 --level 10 --bitmap none --raid-devices 4
/dev/nvme0n1 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme3n1
Test cmd:
fio -name=0 -ioengine=libaio -direct=1 -group_reporting=1
-randseed=2022 -rwmixread=70 -refill_buffers -filename=/dev/md0
-numjobs=16 -runtime=60s -bs=4k -iodepth=256 -rw=randread
Test result:
before this patchset: 2.9 GiB/s
after this patchset: 6.6 Gib/s
Could you please give more details about the test setup, like the
drives used?
test setup is described above, four nvme disks is used.
I was wondering about the model to be able to reproduce it.
Did you use some tools like ftrace to figure out the bottleneck?
Yes, I'm sure the bottleneck is spin_lock(), specifically threads from
multiple nodes try to grab the same lock. By the way, if I bind the
threads to the same node, performance can also improve to 6.6 Gib/s
without this patchset.
Interesting. Maybe you could add all that to the commit message of the
second patch.
Of course, I will do that in next version.
Thanks,
Kuai
Kind regards,
Paul
Please noted that in kunpeng-920, memory access latency is very bad
accross nodes compare to local node, and in other architecture
performance improvement might not be significant.
Yu Kuai (3):
md/raid10: fix improper BUG_ON() in raise_barrier()
md/raid10: convert resync_lock to use seqlock
md/raid10: prevent unnecessary calls to wake_up() in fast path
drivers/md/raid10.c | 88
+++++++++++++++++++++++++++++----------------
drivers/md/raid10.h | 2 +-
2 files changed, 59 insertions(+), 31 deletions(-)
.
