On Fri, Sep 15 2017 at 12:44pm -0400, Ming Lei <ming.lei@xxxxxxxxxx> wrote: > The actual I/O schedule is done in dm-mpath layer, and the > underlying I/O schedule is simply bypassed. > > This patch sets underlying queue's nr_requests as its queue's > queue_depth, then we can get its queue busy feedback by simply > checking if blk_get_request() returns successfully. > > In this way, dm-mpath can reports its queue busy to block layer > effectively, so I/O scheduling is improved much. > > Follows the test results on lpfc*: > > - fio(libaio, bs:4k, dio, queue_depth:64, 64 jobs, > over dm-mpath disk) > - system(12 cores, dual sockets, mem: 64G) > > --------------------------------------- > |v4.13+ |v4.13+ > |+scsi_mq_perf |+scsi_mq_perf+patches > ----------------------------------------- > IOPS(K) |MQ-DEADLINE |MQ-DEADLINE > ------------------------------------------ > read | 30.71 | 343.91 > ----------------------------------------- > randread | 22.98 | 17.17 > ------------------------------------------ > write | 16.45 | 390.88 > ------------------------------------------ > randwrite | 16.21 | 16.09 > --------------------------------------- > > *: > 1) lpfc.lpfc_lun_queue_depth=3, so that it is same with .cmd_per_lun > 2) scsi_mq_perf means the patchset of 'blk-mq-sched: improve SCSI-MQ performance(V4)'[1] > 3) v4.13+: top commit is 46c1e79fee41 Merge branch 'perf-urgent-for-linus' > 4) the patchset 'blk-mq-sched: improve SCSI-MQ performance(V4)' focuses > on improving on SCSI-MQ, and all the test result in that coverletter was > against the raw lpfc/ib(run after 'multipath -F'), instead of dm-mpath. > 5) this patchset itself doesn't depend on the scsi_mq_perf patchset[1] > > [1] https://marc.info/?t=150436555700002&r=1&w=2 > > Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx> Very impressive gains Ming! I'll review in more detail next week but Bart's concerns on patch 1 and 2 definitely need consideration. Thanks, Mike
