On Thu, Oct 12, 2023 at 6:35 AM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote: > > On Thu, Oct 12, 2023 at 1:04 AM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote: > > > > On Wed, Oct 11, 2023 at 8:13 PM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote: > > > > > > On Wed, Oct 11, 2023 at 5:46 AM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote: > > > > > > > > On Tue, Oct 10, 2023 at 6:48 PM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote: > > > > > > > > > > On Tue, Oct 10, 2023 at 5:36 PM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote: > > > > > > > > > > > > On Tue, Oct 10, 2023 at 03:21:47PM -0700, Yosry Ahmed wrote: > > > > > > [...] > > > > > > > > > > > > > > I tried this on a machine with 72 cpus (also ixion), running both > > > > > > > netserver and netperf in /sys/fs/cgroup/a/b/c/d as follows: > > > > > > > # echo "+memory" > /sys/fs/cgroup/cgroup.subtree_control > > > > > > > # mkdir /sys/fs/cgroup/a > > > > > > > # echo "+memory" > /sys/fs/cgroup/a/cgroup.subtree_control > > > > > > > # mkdir /sys/fs/cgroup/a/b > > > > > > > # echo "+memory" > /sys/fs/cgroup/a/b/cgroup.subtree_control > > > > > > > # mkdir /sys/fs/cgroup/a/b/c > > > > > > > # echo "+memory" > /sys/fs/cgroup/a/b/c/cgroup.subtree_control > > > > > > > # mkdir /sys/fs/cgroup/a/b/c/d > > > > > > > # echo 0 > /sys/fs/cgroup/a/b/c/d/cgroup.procs > > > > > > > # ./netserver -6 > > > > > > > > > > > > > > # echo 0 > /sys/fs/cgroup/a/b/c/d/cgroup.procs > > > > > > > # for i in $(seq 10); do ./netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- > > > > > > > -m 10K; done > > > > > > > > > > > > You are missing '&' at the end. Use something like below: > > > > > > > > > > > > #!/bin/bash > > > > > > for i in {1..22} > > > > > > do > > > > > > /data/tmp/netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K & > > > > > > done > > > > > > wait > > > > > > > > > > > > > > > > Oh sorry I missed the fact that you are running instances in parallel, my bad. > > > > > > > > > > So I ran 36 instances on a machine with 72 cpus. I did this 10 times > > > > > and got an average from all instances for all runs to reduce noise: > > > > > > > > > > #!/bin/bash > > > > > > > > > > ITER=10 > > > > > NR_INSTANCES=36 > > > > > > > > > > for i in $(seq $ITER); do > > > > > echo "iteration $i" > > > > > for j in $(seq $NR_INSTANCES); do > > > > > echo "iteration $i" >> "out$j" > > > > > ./netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K >> "out$j" & > > > > > done > > > > > wait > > > > > done > > > > > > > > > > cat out* | grep 540000 | awk '{sum += $5} END {print sum/NR}' > > > > > > > > > > Base: 22169 mbps > > > > > Patched: 21331.9 mbps > > > > > > > > > > The difference is ~3.7% in my runs. I am not sure what's different. > > > > > Perhaps it's the number of runs? > > > > > > > > My base kernel is next-20231009 and I am running experiments with > > > > hyperthreading disabled. > > > > > > Using next-20231009 and a similar 44 core machine with hyperthreading > > > disabled, I ran 22 instances of netperf in parallel and got the > > > following numbers from averaging 20 runs: > > > > > > Base: 33076.5 mbps > > > Patched: 31410.1 mbps > > > > > > That's about 5% diff. I guess the number of iterations helps reduce > > > the noise? I am not sure. > > > > > > Please also keep in mind that in this case all netperf instances are > > > in the same cgroup and at a 4-level depth. I imagine in a practical > > > setup processes would be a little more spread out, which means less > > > common ancestors, so less contended atomic operations. > > > > > > (Resending the reply as I messed up the last one, was not in plain text) > > > > I was curious, so I ran the same testing in a cgroup 2 levels deep > > (i.e /sys/fs/cgroup/a/b), which is a much more common setup in my > > experience. Here are the numbers: > > > > Base: 40198.0 mbps > > Patched: 38629.7 mbps > > > > The regression is reduced to ~3.9%. > > > > What's more interesting is that going from a level 2 cgroup to a level > > 4 cgroup is already a big hit with or without this patch: > > > > Base: 40198.0 -> 33076.5 mbps (~17.7% regression) > > Patched: 38629.7 -> 31410.1 (~18.7% regression) > > > > So going from level 2 to 4 is already a significant regression for > > other reasons (e.g. hierarchical charging). This patch only makes it > > marginally worse. This puts the numbers more into perspective imo than > > comparing values at level 4. What do you think? > > This is weird as we are running the experiments on the same machine. I > will rerun with 2 levels as well. Also can you rerun the page fault > benchmark as well which was showing 9% regression in your original > commit message? Thanks. I will re-run the page_fault tests, but keep in mind that the page fault benchmarks in will-it-scale are highly variable. We run them between kernel versions internally, and I think we ignore any changes below 10% as the benchmark is naturally noisy. I have a couple of runs for page_fault3_scalability showing a 2-3% improvement with this patch :)