On Fri, Oct 20, 2023 at 11:46:36AM +0900, Hyeonggon Yoo wrote: > On Fri, Oct 20, 2023 at 10:27 AM Luiz Capitulino <luizcap@xxxxxxxxxx> wrote: > > > > Hi, > > > > As reported before[1], we found another regression in 6.1 when doing > > performance comparisons with 5.10. This one is caused by CONFIG_DEBUG_PREEMPT > > being enabled by default by the following upstream commit if you have the > > right config dependencies enabled (commit is introduced in v5.16-rc1): > > > > """ > > commit c597bfddc9e9e8a63817252b67c3ca0e544ace26 > > Author: Frederic Weisbecker <frederic@xxxxxxxxxx> > > Date: Tue Sep 14 12:31:34 2021 +0200 > > > > sched: Provide Kconfig support for default dynamic preempt mode > > """ > > > > We found up to 8% performance improvement with CONFIG_DEBUG_PREEMPT > > disabled in different perf benchmarks (including UnixBench process > > creation and redis). The root cause is explained in the commit log > > below which is merged in 6.3 and applies (almost) clealy on 6.1.59. > > Oh, I should've sent it to the stable. Thanks for sending it! Thanks for doing the original fix! :) > Yes, DEBUG_PREEMPT was unintentionally enabled after the introduction > of PREEMPT_DYNAMIC. It was already enabled by default for PREEMPTION=y kernels > but PREEMPT_DYNAMIC always enables PREEMPT_BUILD (and hence PREEMPTION) > so distros that were using PREEMPT_VOLUNTARY are silently affected by that. > > It looks appropriate to be backported to the stable tree (to me). > Hmm but I think it should be backported to 5.15 too? Yeah, I see that Greg applied it to 5.15 and 5.10 as well. I posted it only for 6.1 because the worst case seems to happen after c597bfddc9 where CONFIG_DEBUG_PREEMPT may be enabled automatically. But having the fix in earlier kernels is certainly good as well. - Luiz > > > """ > > commit cc6003916ed46d7a67d91ee32de0f9138047d55f > > Author: Hyeonggon Yoo <42.hyeyoo@xxxxxxxxx> > > Date: Sat Jan 21 12:39:42 2023 +0900 > > > > lib/Kconfig.debug: do not enable DEBUG_PREEMPT by default > > > > In workloads where this_cpu operations are frequently performed, > > enabling DEBUG_PREEMPT may result in significant increase in > > runtime overhead due to frequent invocation of > > __this_cpu_preempt_check() function. > > > > This can be demonstrated through benchmarks such as hackbench where this > > configuration results in a 10% reduction in performance, primarily due to > > the added overhead within memcg charging path. > > """ > > > > [1] https://lore.kernel.org/stable/010edf5a-453d-4c98-9c07-12e75d3f983c@xxxxxxxxxx/
