Hi, On 21/04/21 11:20, Oliver Sang wrote: > hi, Valentin Schneider, > > On Wed, Apr 14, 2021 at 06:17:38PM +0100, Valentin Schneider wrote: >> On 14/04/21 13:21, kernel test robot wrote: >> > Greeting, >> > >> > FYI, we noticed a -13.8% regression of stress-ng.vm-segv.ops_per_sec due to commit: >> > >> > >> > commit: 38ac256d1c3e6b5155071ed7ba87db50a40a4b58 ("[PATCH v5 1/3] sched/fair: Ignore percpu threads for imbalance pulls") >> > url: https://github.com/0day-ci/linux/commits/Valentin-Schneider/sched-fair-load-balance-vs-capacity-margins/20210408-060830 >> > base: https://git.kernel.org/cgit/linux/kernel/git/tip/tip.git 0a2b65c03e9b47493e1442bf9c84badc60d9bffb >> > >> > in testcase: stress-ng >> > on test machine: 96 threads Intel(R) Xeon(R) Gold 6252 CPU @ 2.10GHz with 192G memory >> > with following parameters: >> > >> > nr_threads: 10% >> > disk: 1HDD >> > testtime: 60s >> > fs: ext4 >> > class: os >> > test: vm-segv >> > cpufreq_governor: performance >> > ucode: 0x5003006 >> > >> > >> >> That's almost exactly the same result as [1], which is somewhat annoying >> for me because I wasn't able to reproduce those results back then. Save >> from scrounging the exact same machine to try this out, I'm not sure what's >> the best way forward. I guess I can re-run the workload on whatever >> machines I have and try to spot any potentially problematic pattern in the >> trace... > > what's the machine model you used upon which the regression cannot be reproduced? > we could check if we have similar model then re-check on the our machine. > I tested this on: o Ampere eMAG (arm64, 32 cores) o 2-socket Xeon E5-2690 (x86, 40 cores) and found at worse a -0.3% regression and at best a 2% improvement. I know that x86 box is somewhat ancient, but it's been my go-to "have I broken x86?" test victim for a while :-) > BTW, we supplied perf data in original report, not sure if they are helpful? > or do you have suggestion which kind of data will be more helpful to you? > we will continuously improve our report based on suggestions from community. > Thanks a lot! > Staring at it some more, I notice a huge uptick in: - major page faults (+315.2% and +270%) - cache misses (+125.2% and +131.0%) I don't really get the page faults; the cache misses I could somewhat understand: this is adding p->flags and (p->set_child_tid)->flags accesses, which are in different cachelines than p->se and p->cpus_mask used in can_migrate_task(). I think I could dig some more into this with perf, but I'd need to be able to reproduce this locally first... >> >> [1]: http://lore.kernel.org/r/20210223023004.GB25487@xsang-OptiPlex-9020