Re: [PATCH] sched/fair: Let sync wakeups target the waker's core

Hi Vineeth,

I tested the things you wanted on the same box as before in the current 
baseline (d58772d8520c) and baseline + patch. 

command: perf bench sched pipe

baseline:
run 1: usecs/op: 1.435285, ops/sec: 696725
run 2: usecs/op: 1.305019, ops/sec: 765685
run 3: usecs/op: 1.143834, ops/sec: 874176
run 4: usecs/op: 1.549996, ops/sec: 645162
run 5: usecs/op: 1.136511, ops/sec: 879885

baseline + patch:
run 1: usecs/op: 1.362729, ops/sec: 733821
run 2: usecs/op: 1.420710, ops/sec: 703873
run 3: usecs/op: 1.229553, ops/sec: 813303
run 4: usecs/op: 1.453649, ops/sec: 687923
run 5: usecs/op: 1.204788, ops/sec: 839921

I don't think these numbers really mean anything.
Range is too big, as an example, on baseline the 
biggest usecs/op is 1.54996, while the smallest one 
is 1.136511. And this pattern just repeats itself.
The range is smaller on the baseline + patch,
but I don't really know what it means.

command: perf stat -r 10 -e cycles,cache-misses,L1-dcache-loads,L1-dcache-load-misses,cpu/event=0x64,umask=0x08/ 
perf bench sched pipe

I got the L1-dcache-load-misses and L1-dcache-loads from the perf list, as for the other one I got it from
tools/perf/pmu-events/arch/x86/amdzen3/recommended.json, l2_cache_misses_from_dc_misses.

I ran this once on both the baseline and baseline + patch since it runs 10 times with -r.

baseline: 

cycles: 12317680674
cache-misses: 19567069
L1-dcache-loads: 2792471451
L1-dcache-load-misses: 52924092
L2(event = 0x64, umask = 0x08): 18696435

baseline + patch:

cycles: 12304183071
cache-misses: 20105690
L1-dcache-loads: 2796772374
L1-dcache-load-misses: 50928594
L2(event = 0x64, umask = 0x08): 18928237

In the old test, the cycles decreased 1.40%, but in this one the cycles decreased 0.11%. Also
cache-misses increased 4.60% in the old test, but in this one it increased 2.75%.

I think this may be because the baseline is different, but that is just a guess.


I also tested LLC, or L3, same as L2 I got the values from /amdzen3/recommended.json, l3_misses.

baseline: 

L3(event=0x04, umask=0x01): 23064400

baseline + patch:

L3(event=0x04, umask=0x01): 22870280

L1-dcache-load-misses went down 3.77% while L2 went up 1.24% and
L3 went down 0.84%.

Also please drop the Tested-by, I'll retest on v2 and send a new one. 

Thanks, 
Kayra