On Wed, Dec 01, 2021 at 01:03:48AM +0900, Alexey Avramov wrote: > I tested this [1] patch on top of 5.16-rc2. It's the same test with 10 tails. > > - with noswap > > Summary: > > 2021-11-30 23:32:36,890: Stall times for the last 548.6s: > 2021-11-30 23:32:36,890: ----------- > 2021-11-30 23:32:36,891: some cpu 3.7s, avg 0.7% > 2021-11-30 23:32:36,891: ----------- > 2021-11-30 23:32:36,891: some io 187.6s, avg 34.2% > 2021-11-30 23:32:36,891: full io 178.3s, avg 32.5% > 2021-11-30 23:32:36,891: ----------- > 2021-11-30 23:32:36,892: some memory 392.2s, avg 71.5% > 2021-11-30 23:32:36,892: full memory 390.7s, avg 71.2% > > full psi: > https://raw.githubusercontent.com/hakavlad/cache-tests/main/516-reclaim-throttle/516-rc2/patch5/noswap/psi > > mem: > https://raw.githubusercontent.com/hakavlad/cache-tests/main/516-reclaim-throttle/516-rc2/patch5/noswap/mem > Ok, taking just noswap in isolation, this is what I saw when running firefox + youtube vido and running tail /dev/zero 10 times in a row 2021-11-30 17:10:11,817: ================================= 2021-11-30 17:10:11,817: Peak values: avg10 avg60 avg300 2021-11-30 17:10:11,817: ----------- ------ ------ ------ 2021-11-30 17:10:11,817: some cpu 1.00 0.96 0.56 2021-11-30 17:10:11,817: ----------- ------ ------ ------ 2021-11-30 17:10:11,817: some io 0.24 0.06 0.04 2021-11-30 17:10:11,817: full io 0.24 0.06 0.01 2021-11-30 17:10:11,817: ----------- ------ ------ ------ 2021-11-30 17:10:11,817: some memory 2.48 0.51 0.38 2021-11-30 17:10:11,817: full memory 2.48 0.51 0.37 2021-11-30 17:10:11,817: ================================= 2021-11-30 17:10:11,817: Stall times for the last 53.7s: 2021-11-30 17:10:11,817: ----------- 2021-11-30 17:10:11,817: some cpu 0.4s, avg 0.8% 2021-11-30 17:10:11,817: ----------- 2021-11-30 17:10:11,817: some io 0.1s, avg 0.2% 2021-11-30 17:10:11,817: full io 0.1s, avg 0.2% 2021-11-30 17:10:11,817: ----------- 2021-11-30 17:10:11,817: some memory 0.3s, avg 0.6% 2021-11-30 17:10:11,817: full memory 0.3s, avg 0.6% Obviously a fairly different experience and most likely due to the underlying storage. Can you run the same test but after doing this $ echo 1 > /sys/kernel/debug/tracing/events/vmscan/mm_vmscan_throttled/enable $ cat /sys/kernel/debug/tracing/trace_pipe > trace.out and send me the trace.out file please? -- Mel Gorman SUSE Labs
