> -----Original Message----- > From: linux-scsi-owner@xxxxxxxxxxxxxxx [mailto:linux-scsi- > owner@xxxxxxxxxxxxxxx] On Behalf Of Douglas Gilbert > Sent: Wednesday, 18 June, 2014 8:54 PM > To: SCSI development list > Subject: [RFC] scsi_debug: locks and delays > > Currently the scsi_debug driver wraps every queued command > with the host_lock and every mid-level completion callback > (apart from delay=0) with a spinlock. > > Attached is a patch against Linus's tree that also applies > to lk 3.15.1 . It attempts to address some of these issues. > > ChangeLog > - 'host_lock' option added that simply drops that lock > when host_lock=0 (which is the default) > - allow delay=-1 [delay=1 (jiffy) is still the default] > that uses a tasklet to schedule the response quickly > - for completions (when delay!=0) the callback to the > mid-level is un-(spin)locked > - completions are counted; can be viewed with > cat /proc/scsi/scsi_debug/<host_num> > - delay_override removed from TEST UNIT READY. > This makes 'sg_turs -n 1m -t /dev/bsg/<hctl>' a more > realistic test of command overhead. I get about 100k > IOPS on my laptop. > > This patch has been lightly tested. Perhaps someone could > throw a scsi-mq test at it. > > Comments welcome. > > Doug Gilbert This is indeed much better with scsi-mq than without scsi-mq, which was exhibiting 90K to 171K IOPS on various systems when I last tried it. Using one scsi_debug device in fake_rw mode, 4 KiB random reads get these results: 236K IOPS 1 thread 1081K IOPS 6 threads 1205K IOPS 6 threads, iodepth_batch=4 iodepth_batch_complete=4 1270K IOPS 6 threads, iodepth_batch=16 iodepth_batch_complete=16 Toying with a few sysfs properties: * add_random=1 drops from 1270K to 1170K IOPS. * nomerges=0, 1, or 2 doesn't matter. The average queue size never exceeds 3 because they complete so fast, so there's not much searching going on. I created the device with: modprobe scsi_debug fake_rw=1 delay=0 physblk_exp=3 dev_size_mb=128 fio options (adjust as needed): $ cat 1drive_sdah.fio [global] direct=1 ioengine=libaio norandommap randrepeat=0 bs=4096 iodepth=96 numjobs=1 #numjobs=6 runtime=216000 time_based=1 group_reporting thread gtod_reduce=1 #iodepth_batch=4 #iodepth_batch_complete=4 cpus_allowed=0-5 cpus_allowed_policy=split rw=randread [4_KiB_RR_drive_ah] filename=/dev/sdah --- Rob Elliott HP Server Storage -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
