On 16/03/18 19:10, Evgeniy Didin wrote: > Hello Adrian, > >> Yes. Unfortunately the clock used is not accurate enough to correctly order >> the events across different CPUs, which makes it very hard to see delays >> between requests. You could try a different clock - refer the --clockid >> option to perf record. >> >> Nevertheless it shows there are no I/O errors which means the error recovery >> can be ruled out as a problem. >> >> The issue could be caused by the I/O scheduler. Under blk-mq the default >> scheduler is the mq-deadline scheduler whereas without blk-mq you would >> probably have been using cfq by default. You could try the bfq scheduler: >> >> echo bfq > /sys/block/mmcblk0/queue/scheduler >> >> But you might need to add it to the kernel config i.e. >> >> CONFIG_IOSCHED_BFQ=y >> > Switching from mq-deadline scheduler to bfq fixed the issue. > Also bonnie++ results have changed: > -----------------------------------------------<8---------------------------------------------------------------------------- > bfq scheduler: > ARCLinux,512M,6463,87,7297,0,5450,0,9827,99,342952,99,+++++,+++,16,17525,100,+++++,+++,24329,99,17621,100,+++++,+++,24001,101 > > mq-deadline scheduler: > ARCLinux,512M,4453,36,6474,1,5852,0,12940,99,344329,100,+++++,+++,16,22168,98,+++++,+++,32760,99,22755,100,+++++,+++,32205,100 > -----------------------------------------------<8---------------------------------------------------------------------------- > As I see, the performance of sequential input per char and of file > operations have decreased for ~25%. You may need to aggregate more runs, and also compare to BFQ with blk-mq against CFQ without blk-mq. If you think BFQ is under-performing, then contact the BFQ maintainers. > > Do you have any idea what could be a reason for such a long stalling in > case of mq-deadline IOscheduler? Write starvation. > I would expect if there is some long > async operation, kernel should not be blocked. The kernel is not blocked. AFAICT it is the EXT4 journal that is blocked waiting on a write. > But what we see using > mq-deadline is kernel blocked in bit_wait_io(). Do you think this is a > valid behavior at least in case of mq-deadline IOscheduler? mq-deadline is designed to favour reads over writes, so in that sense some amount of write-starvation is normal. > >> Alternatively you could fiddle with the scheduler parameters: >> >> With mq-deadline they are: >> >> # grep -H . /sys/block/mmcblk0/queue/iosched/* >> /sys/block/mmcblk0/queue/iosched/fifo_batch:16 >> /sys/block/mmcblk0/queue/iosched/front_merges:1 >> /sys/block/mmcblk0/queue/iosched/read_expire:500 >> /sys/block/mmcblk0/queue/iosched/write_expire:5000 >> /sys/block/mmcblk0/queue/iosched/writes_starved:2 >> >> You could try decreasing the write_expire and/or fifo_batch. > It seems that decreasing doesn't affect on this issue. That is surprising. You could also try writes_starved=1 or writes_starved=0.
