Feng Tang <feng.tang@xxxxxxxxx> writes: > When doing performance tuning or debugging performance regressions, > more and more cases are found to be related to false sharing [1][2][3], > and the situation can be worse for newer platforms with hundreds of > CPUs. There are already many commits in current kernel specially > for mitigating the performance degradation due to false sharing. > > False sharing could harm the performance silently without being > noticed, due to reasons like: > * data members of a big data structure randomly sitting together > in one cache line > * global data of small size are linked compactly together > > So it's better to make a simple document about the normal pattern > of false sharing, basic ways to mitigate it and call out to > developers to pay attention during code-writing. > > [ Many thanks to Dave Hansen, Ying Huang, Tim Chen, Julie Du and > Yu Chen for their contributions ] > > [1]. https://lore.kernel.org/lkml/20220619150456.GB34471@xsang-OptiPlex-9020/ > [2]. https://lore.kernel.org/lkml/20201102091543.GM31092@shao2-debian/ > [3]. https://lore.kernel.org/lkml/20230307125538.818862491@xxxxxxxxxxxxx/ > > Signed-off-by: Feng Tang <feng.tang@xxxxxxxxx> > Reviewed-by: Randy Dunlap <rdunlap@xxxxxxxxxxxxx> > Reviewed-by: Shakeel Butt <shakeelb@xxxxxxxxxx> Applied, thanks. jon
