[lkml]   [2014]   [Jun]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: scsi-mq
On 06/18/14 05:44, Jens Axboe wrote:
> Thanks for posting these numbers, Bart. The CPU utilization and IOPS
> speak a very clear message. The only mystery is why the singe threaded
> performance is down. That we need to get sort, but it's not a show
> stopper for inclusion.
> If you run the single threaded tests and watch for queue depths, is
> there a difference between blk-mq=y/scsi-mq and the stock kernel?

Hello Jens,

Fio reports the same queue depth for use_blk_mq=Y (mq below) and
use_blk_mq=N (sq below), namely ">=64". However, the number of context
switches differs significantly for the random read-write tests. From
the fio output for these tests:

$ grep ctx= {sq,mq}/randrw*
sq/randrw-1.txt: cpu : usr=10.25%, sys=89.42%, ctx=2210, majf=0, minf=1
sq/randrw-2.txt: cpu : usr=10.23%, sys=89.34%, ctx=5003, majf=0, minf=2
sq/randrw-3.txt: cpu : usr=9.36%, sys=90.21%, ctx=7947, majf=0, minf=3
sq/randrw-4.txt: cpu : usr=8.96%, sys=90.10%, ctx=19308, majf=0, minf=4
sq/randrw-5.txt: cpu : usr=8.97%, sys=89.70%, ctx=31494, majf=0, minf=5
sq/randrw-6.txt: cpu : usr=8.39%, sys=90.08%, ctx=47826, majf=0, minf=6
sq/randrw-7.txt: cpu : usr=7.65%, sys=89.65%, ctx=130563, majf=0, minf=7
sq/randrw-8.txt: cpu : usr=6.47%, sys=84.08%, ctx=753140, majf=0, minf=8
mq/randrw-1.txt: cpu : usr=1.43%, sys=14.43%, ctx=500998, majf=0, minf=1
mq/randrw-2.txt: cpu : usr=1.37%, sys=14.13%, ctx=979842, majf=0, minf=2
mq/randrw-3.txt: cpu : usr=1.47%, sys=14.81%, ctx=1547996, majf=0, minf=3
mq/randrw-4.txt: cpu : usr=1.79%, sys=16.51%, ctx=2321154, majf=0, minf=4
mq/randrw-5.txt: cpu : usr=2.49%, sys=22.09%, ctx=4145747, majf=0, minf=5
mq/randrw-6.txt: cpu : usr=2.98%, sys=27.07%, ctx=6356183, majf=0, minf=6
mq/randrw-7.txt: cpu : usr=3.39%, sys=30.48%, ctx=8675960, majf=0, minf=7
mq/randrw-8.txt: cpu : usr=3.37%, sys=31.46%, ctx=10462001, majf=0, minf=8

It seems like with the traditional SCSI mid-layer and block core (sq)
that the number of context switches does not depend too much on the
number of I/O operations but that for the multi-queue SCSI core there
are a little bit more than two context switches per I/O in the
particular test I ran. The "randrw" script I used for this test takes
SCSI LUNs as arguments (/dev/sdX) and starts the fio tool as follows:

for d in "$@"; do
if [ ! -e "$d" ]; then
echo "Error: device $d not found."
exit 1
bdev="/sys/class/block/$(basename $d)"
if [ -e $bdev/queue ]; then
echo 0 >$bdev/queue/add_random
echo 0 >$bdev/queue/rotational
echo 2 >$bdev/queue/rq_affinity
echo noop >$bdev/queue/scheduler

"$(dirname $0)"/disable-frequency-scaling

fio --bs=512 --ioengine=libaio --rw=randrw --iodepth=128 \
--iodepth_batch=64 --iodepth_batch_complete=64 \
--buffered=0 --norandommap --thread --loops=$((2**31)) \
--runtime=60 --group_reporting --gtod_reduce=1 --invalidate=1 \
$(for d in "$@"; do echo --name=$d --filename=$d; done)

"$(dirname $0)"/restore-frequency-scaling


 \ /
  Last update: 2014-06-18 10:01    [W:0.096 / U:0.348 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site