Hi Andrew and Mel, While testing per-vma locks on mm-unstable branch I noticed a sizable regression when running pft-threads test from Mel's mmtests suite: BASE MM_UNSTABLE Hmean faults/sec-1 469201.7282 ( 0.00%) 305002.3892 * -35.00%* Hmean faults/sec-4 1754465.6221 ( 0.00%) 1183736.3124 * -32.53%* Hmean faults/sec-7 2808141.6711 ( 0.00%) 1964469.4590 * -30.04%* Hmean faults/sec-12 3750307.7553 ( 0.00%) 2811359.5393 * -25.04%* Hmean faults/sec-21 4145672.4677 ( 0.00%) 3082305.9582 * -25.65%* Hmean faults/sec-30 3775722.5726 ( 0.00%) 2918088.2083 * -22.71%* Hmean faults/sec-48 4152563.5864 ( 0.00%) 3219418.0414 * -22.47%* Hmean faults/sec-56 4163868.7111 ( 0.00%) 3124612.3198 * -24.96%* This is comparing mm-unstable ToT (c04e06325f11 "mm: multi-gen LRU: simplify lru_gen_look_around()") with 6.2-rc3, which was my previous baseline (both are tested without my patchset). Not sure if this is expected, so sending a warning. If this is unexpected, it would also be great if someone can confirm this is not a fluke on my side. The test machine I use is a NUMA 2-socket Intel Xeon E5-2690-v4 (Broadwell microarchitecture). Command I used to reproduce is: run-mmtests.sh --no-monitor --config configs/config-workload-pft-threads <name> Thanks, Suren.
