On Thu, Oct 17, 2024 at 10:58:38AM +0800, Oliver Sang wrote: > hi, Lorenzo, > > On Tue, Oct 15, 2024 at 08:56:28PM +0100, Lorenzo Stoakes wrote: > > On Fri, Oct 11, 2024 at 08:26:37AM +0100, Lorenzo Stoakes wrote: > > [snip] > > > > > Thanks for testing this suffices to rule this one out... I will try to get a > > > functional and reliable performance environment locally so I can properly > > > address this and then we can try something else. > > > > > > Thanks! > > > Lorenzo > > > > > > > OK Oliver, could you try the below patch? I have got aim9.brk up and > > running locally and for me this seems to address the issue. > > > > This is against Andrew's tree [0] in the mm-unstable branch. It should > > hopefully apply cleanly to -next also. > > I found the patch still be able to applied to cacded5e42 cleanly, so below data > still based on this applyment. > > $ git log --oneline 9cecc5dc893886 > 9cecc5dc893886 mm: add expand-only VMA merge mode and optimise do_brk_flags() > cacded5e42b960 mm: avoid using vma_merge() for new VMAs > fc21959f74bc11 mm: abstract vma_expand() to use vma_merge_struct > ... > > again, if some patches in mm-unstable or -next have some impacts, please let me > know then I can re-apply the patch and do the tests again. thanks > > > by this patch, we do see performance recovery but not fully. > > e.g. for > model: Granite Rapids > nr_node: 1 > nr_cpu: 240 > memory: 192G > > we got better score from the patch than cacded5e42b960, but still 2.0% > regression than fc21959f74bc11 (the parent of cacded5e42b960) Thanks for this. As far as I'm concerned this puts us into noise territory, so we'll go with this as the solution! A side-note, the brk2 test from the will-it-scale suite, written explicitly to be more real-world representative, sees an actual performance _improvement_ here (though small). So overall I'm comfortable with this, we can revisit if anybody raises any objection! The benefit in de-duplicating code is very significant. Thanks for all your help, hugely appreciated! Cheers, Lorenzo [snip]
