On 30/09/2017 11:27, Ming Lei wrote:
Hi Jens,
In Red Hat internal storage test wrt. blk-mq scheduler, we
found that I/O performance is much bad with mq-deadline, especially
about sequential I/O on some multi-queue SCSI devcies(lpfc, qla2xxx,
SRP...)
Turns out one big issue causes the performance regression: requests
are still dequeued from sw queue/scheduler queue even when ldd's
queue is busy, so I/O merge becomes quite difficult to make, then
sequential IO degrades a lot.
This issue becomes one of mains reasons for reverting default SCSI_MQ
in V4.13.
The 1st patch takes direct issue in blk_mq_request_bypass_insert(),
then we can improve dm-mpath's performance in part 2, which will
be posted out soon.
The 2nd six patches improve this situation, and brings back
some performance loss.
With this change, SCSI-MQ sequential I/O performance is
improved much, Paolo reported that mq-deadline performance
improved much[2] in his dbench test wrt V2. Also performanc
improvement on lpfc/qla2xx was observed with V1.[1]
Please consider it for V4.15.
[1] http://marc.info/?l=linux-block&m=150151989915776&w=2
[2] https://marc.info/?l=linux-block&m=150217980602843&w=2
I tested this series for the SAS controller on HiSilicon hip07 platform
as I am interested in enabling MQ for this driver. Driver is
./drivers/scsi/hisi_sas/.
So I found that that performance is improved when enabling default
SCSI_MQ with this series vs baseline. However, it is still not as a good
as when default SCSI_MQ is disabled.
Here are some figures I got with fio:
4.14-rc2 without default SCSI_MQ
read, rw, write IOPS
952K, 133K/133K, 800K
4.14-rc2 with default SCSI_MQ
read, rw, write IOPS
311K, 117K/117K, 320K
This series* without default SCSI_MQ
read, rw, write IOPS
975K, 132K/132K, 790K
This series* with default SCSI_MQ
read, rw, write IOPS
770K, 164K/164K, 594K
Please note that hisi_sas driver does not enable mq by exposing multiple
queues to upper layer (even though it has multiple queues). I have been
playing with enabling it, but my performance is always worse...
* I'm using
https://github.com/ming1/linux/commits/blk_mq_improve_scsi_mpath_perf_V5.1,
as advised by Ming Lei.
Thanks,
John
V5:
- address some comments from Omar
- add Tested-by & Reveiewed-by tag
- use direct issue for blk_mq_request_bypass_insert(), and
start to consider to improve sequential I/O for dm-mpath
- only include part 1(the original patch 1 ~ 6), as suggested
by Omar
V4:
- add Reviewed-by tag
- some trival change: typo fix in commit log or comment,
variable name, no actual functional change
V3:
- totally round robin for picking req from ctx, as suggested
by Bart
- remove one local variable in __sbitmap_for_each_set()
- drop patches of single dispatch list, which can improve
performance on mq-deadline, but cause a bit degrade on
none because all hctxs need to be checked after ->dispatch
is flushed. Will post it again once it is mature.
- rebase on v4.13-rc6 with block for-next
V2:
- dequeue request from sw queues in round roubin's style
as suggested by Bart, and introduces one helper in sbitmap
for this purpose
- improve bio merge via hash table from sw queue
- add comments about using DISPATCH_BUSY state in lockless way,
simplifying handling on busy state,
- hold ctx->lock when clearing ctx busy bit as suggested
by Bart
Ming Lei (7):
blk-mq: issue rq directly in blk_mq_request_bypass_insert()
blk-mq-sched: fix scheduler bad performance
sbitmap: introduce __sbitmap_for_each_set()
blk-mq: introduce blk_mq_dequeue_from_ctx()
blk-mq-sched: move actual dispatching into one helper
blk-mq-sched: improve dispatching from sw queue
blk-mq-sched: don't dequeue request until all in ->dispatch are
flushed
block/blk-core.c | 3 +-
block/blk-mq-debugfs.c | 1 +
block/blk-mq-sched.c | 104 ++++++++++++++++++++++++++++++++++++-------
block/blk-mq.c | 114 +++++++++++++++++++++++++++++++++++++++---------
block/blk-mq.h | 4 +-
drivers/md/dm-rq.c | 2 +-
include/linux/blk-mq.h | 3 ++
include/linux/sbitmap.h | 64 +++++++++++++++++++--------
8 files changed, 238 insertions(+), 57 deletions(-)
