On Mon, Aug 24, 2020 at 01:24:20PM -0700, Hugh Dickins wrote: > On Mon, 24 Aug 2020, Andrew Morton wrote: > > On Mon, 24 Aug 2020 20:54:33 +0800 Alex Shi <alex....@linux.alibaba.com> > > wrote: > Andrew demurred on version 17 for lack of review. Alexander Duyck has > been doing a lot on that front since then. I have intended to do so, > but it's a mirage that moves away from me as I move towards it: I have
Same, I haven't been able to keep up with the versions or the recent review feedback. I got through about half of v17 last week and hope to have more time for the rest this week and beyond. > > > Following Daniel Jordan's suggestion, I have run 208 'dd' with on 104 > > > containers on a 2s * 26cores * HT box with a modefied case: Alex, do you have a pointer to the modified readtwice case? Even better would be a description of the problem you're having in production with lru_lock. We might be able to create at least a simulation of it to show what the expected improvement of your real workload is. > > > https://git.kernel.org/pub/scm/linux/kernel/git/wfg/vm-scalability.git/tree/case-lru-file-readtwice > > > With this patchset, the readtwice performance increased about 80% > > > in concurrent containers. > > > > That's rather a slight amount of performance testing for a huge > > performance patchset! > > Indeed. And I see that clause about readtwice performance increased 80% > going back eight months to v6: a lot of fundamental bugs have been fixed > in it since then, so I do think it needs refreshing. It could be faster > now: v16 or v17 fixed the last bug I knew of, which had been slowing > down reclaim considerably. > > When I last timed my repetitive swapping loads (not loads anyone sensible > would be running with), across only two memcgs, Alex's patchset was > slightly faster than without: it really did make a difference. But > I tend to think that for all patchsets, there exists at least one > test that shows it faster, and another that shows it slower. > > > Is more detailed testing planned? > > Not by me, performance testing is not something I trust myself with, > just get lost in the numbers: Alex, this is what we hoped for months > ago, please make a more convincing case, I hope Daniel and others > can make more suggestions. But my own evidence suggests it's good. I ran a few benchmarks on v17 last week (sysbench oltp readonly, kerndevel from mmtests, a memcg-ized version of the readtwice case I cooked up) and then today discovered there's a chance I wasn't running the right kernels, so I'm redoing them on v18. Plan to look into what other, more "macro" tests would be sensitive to these changes.