Re: [tip: sched/core] sched/fair: Only update stats for allowed CPUs when looking for dst group

[Erwan Velu]

Hi Adam, Peter,

I've been chasing a performance bug on my systems and found that this
commit was a
real game-changer on my production server.

I'm using dual AMD Turin 9845 (160 cores each), with the following
NUMA topology:
Socket 0:
  NUMA node0 CPU(s):         0-15,320-335
  NUMA node1 CPU(s):         16-31,336-351
  NUMA node2 CPU(s):         32-47,352-367
  NUMA node3 CPU(s):         48-63,368-383
  NUMA node4 CPU(s):         64-79,384-399
  NUMA node5 CPU(s):         80-95,400-415
  NUMA node6 CPU(s):         96-111,416-431
  NUMA node7 CPU(s):         112-127,432-447
  NUMA node8 CPU(s):         128-143,448-463
  NUMA node9 CPU(s):         144-159,464-479

Socket 1:
  NUMA node10 CPU(s):        160-175,480-495
  NUMA node11 CPU(s):        176-191,496-511
  NUMA node12 CPU(s):        192-207,512-527
  NUMA node13 CPU(s):        208-223,528-543
  NUMA node14 CPU(s):        224-239,544-559
  NUMA node15 CPU(s):        240-255,560-575
  NUMA node16 CPU(s):        256-271,576-591
  NUMA node17 CPU(s):        272-287,592-607
  NUMA node18 CPU(s):        288-303,608-623
  NUMA node19 CPU(s):        304-319,624-639

I was able to setup a reproducer by using the wonderful stress-ng tool
like the following :
    [root@host]$ stress-ng --taskset 0-175 --timeout 30 --metrics
--cpu-method float128 --cpu 176

On my 6.18.38 and 6.12.95, the latest LTS at the date, I have very
weird performance like :
    stress-ng: info:  [11329] dispatching hogs: 176 cpu
    stress-ng: metrc: [11329] stressor       bogo ops real time  usr
time  sys time   bogo ops/s     bogo ops/s CPU used per       RSS Max
    stress-ng: metrc: [11329]                           (secs)
(secs)    (secs)   (real time) (usr+sys time) instance (%)
(KB)
    stress-ng: metrc: [11329] cpu             5515290     29.97
1311.82      0.02    184034.19        4204.23        24.87
0
    stress-ng: info:  [11329] skipped: 0

/!\ Stress-ng reports 25% CPU usage, resulting in only 5M bogo-ops.

When applying this patch on top of the two LTS releases, I get the
following result:
    stress-ng: info:  [11329] dispatching hogs: 176 cpu
    stress-ng: metrc: [11329] stressor       bogo ops real time  usr
time  sys time   bogo ops/s     bogo ops/s CPU used per       RSS Max
    stress-ng: metrc: [11329]                           (secs)
(secs)    (secs)   (real time) (usr+sys time) instance (%)
(KB)
    stress-ng: metrc: [11329] cpu            19306125     29.99
5236.45      0.06    643718.82        3686.83        99.20
0
    stress-ng: info:  [11329] skipped: 0

The CPU usage gets an expected 99% and 19M+ bogo-ops.
That's a massive 4x performance impact on a CPU-bound application.

This outcome is 100% reproducible and particularly visible with this
stress-ng reproducer, but
also exists in other conditions when choosing other NUMA nodes and CPUs.

Considering the impact of this change on performance, can we request a
backport to all stable kernels?
Erwan,