On 3/27/22 19:20, Bart Van Assche wrote:
Thanks, this helps a lot. can_queue = 1 as I was suspecting. In a quick test I noticed that I/O is about 10x slower for queue depth 1 and the v5.17 mq-deadline scheduler. I will take a closer look at this tomorrow.
(replying to my own email) Please take a look at the two new tests in https://github.com/osandov/blktests/pull/87. The results of that test for kernel v5.13 show that enabling an I/O scheduler speeds up the I/O workload except when using BFQ (times are in centiseconds): # cat /home/bart/software/blktests/results/nodev/block/032.full bfq: 465 vs 452: pass kyber: 243 vs 452: pass mq-deadline: 230 vs 452: pass The results for kernel v5.16 shows that enabling an I/O scheduler slows down the I/O workload up to 2.2x: # cat /home/bart/software/blktests/results/nodev/block/032.full bfq: 920 vs 419: fail kyber: 732 vs 419: fail mq-deadline: 751 vs 419: fail In other words, this is not an mq-deadline issue. Bart.
