On Thu, Oct 12, 2023 at 2:39 PM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote: > > On Thu, Oct 12, 2023 at 2:20 PM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote: > > > [...] > > > > > > Yes this looks better. I think we should also ask intel perf and > > > phoronix folks to run their benchmarks as well (but no need to block > > > on them). > > > > Anything I need to do for this to happen? (I thought such testing is > > already done on linux-next) > > Just Cced the relevant folks. > > Michael, Oliver & Feng, if you have some time/resource available, > please do trigger your performance benchmarks on the following series > (but nothing urgent): > > https://lore.kernel.org/all/20231010032117.1577496-1-yosryahmed@xxxxxxxxxx/ Thanks for that. > > > > > Also, any further comments on the patch (or the series in general)? If > > not, I can send a new commit message for this patch in-place. > > Sorry, I haven't taken a look yet but will try in a week or so. Sounds good, thanks. Meanwhile, Andrew, could you please replace the commit log of this patch as follows for more updated testing info: Subject: [PATCH v2 3/5] mm: memcg: make stats flushing threshold per-memcg A global counter for the magnitude of memcg stats update is maintained on the memcg side to avoid invoking rstat flushes when the pending updates are not significant. This avoids unnecessary flushes, which are not very cheap even if there isn't a lot of stats to flush. It also avoids unnecessary lock contention on the underlying global rstat lock. Make this threshold per-memcg. The scheme is followed where percpu (now also per-memcg) counters are incremented in the update path, and only propagated to per-memcg atomics when they exceed a certain threshold. This provides two benefits: (a) On large machines with a lot of memcgs, the global threshold can be reached relatively fast, so guarding the underlying lock becomes less effective. Making the threshold per-memcg avoids this. (b) Having a global threshold makes it hard to do subtree flushes, as we cannot reset the global counter except for a full flush. Per-memcg counters removes this as a blocker from doing subtree flushes, which helps avoid unnecessary work when the stats of a small subtree are needed. Nothing is free, of course. This comes at a cost: (a) A new per-cpu counter per memcg, consuming NR_CPUS * NR_MEMCGS * 4 bytes. The extra memory usage is insigificant. (b) More work on the update side, although in the common case it will only be percpu counter updates. The amount of work scales with the number of ancestors (i.e. tree depth). This is not a new concept, adding a cgroup to the rstat tree involves a parent loop, so is charging. Testing results below show no significant regressions. (c) The error margin in the stats for the system as a whole increases from NR_CPUS * MEMCG_CHARGE_BATCH to NR_CPUS * MEMCG_CHARGE_BATCH * NR_MEMCGS. This is probably fine because we have a similar per-memcg error in charges coming from percpu stocks, and we have a periodic flusher that makes sure we always flush all the stats every 2s anyway. This patch was tested to make sure no significant regressions are introduced on the update path as follows. The following benchmarks were ran in a cgroup that is 2 levels deep (/sys/fs/cgroup/a/b/): (1) Running 22 instances of netperf on a 44 cpu machine with hyperthreading disabled. All instances are run in a level 2 cgroup, as well as netserver: # netserver -6 # netperf -6 -H ::1 -l 60 -t TCP_SENDFILE -- -m 10K Averaging 20 runs, the numbers are as follows: Base: 40198.0 mbps Patched: 38629.7 mbps (-3.9%) The regression is minimal, especially for 22 instances in the same cgroup sharing all ancestors (so updating the same atomics). (2) will-it-scale page_fault tests. These tests (specifically per_process_ops in page_fault3 test) detected a 25.9% regression before for a change in the stats update path [1]. These are the numbers from 10 runs (+ is good) on a machine with 256 cpus: LABEL | MEAN | MEDIAN | STDDEV | ------------------------------+-------------+-------------+------------- page_fault1_per_process_ops | | | | (A) base | 270249.164 | 265437.000 | 13451.836 | (B) patched | 261368.709 | 255725.000 | 13394.767 | | -3.29% | -3.66% | | page_fault1_per_thread_ops | | | | (A) base | 242111.345 | 239737.000 | 10026.031 | (B) patched | 237057.109 | 235305.000 | 9769.687 | | -2.09% | -1.85% | | page_fault1_scalability | | | (A) base | 0.034387 | 0.035168 | 0.0018283 | (B) patched | 0.033988 | 0.034573 | 0.0018056 | | -1.16% | -1.69% | | page_fault2_per_process_ops | | | (A) base | 203561.836 | 203301.000 | 2550.764 | (B) patched | 197195.945 | 197746.000 | 2264.263 | | -3.13% | -2.73% | | page_fault2_per_thread_ops | | | (A) base | 171046.473 | 170776.000 | 1509.679 | (B) patched | 166626.327 | 166406.000 | 768.753 | | -2.58% | -2.56% | | page_fault2_scalability | | | (A) base | 0.054026 | 0.053821 | 0.00062121 | (B) patched | 0.053329 | 0.05306 | 0.00048394 | | -1.29% | -1.41% | | page_fault3_per_process_ops | | | (A) base | 1295807.782 | 1297550.000 | 5907.585 | (B) patched | 1275579.873 | 1273359.000 | 8759.160 | | -1.56% | -1.86% | | page_fault3_per_thread_ops | | | (A) base | 391234.164 | 390860.000 | 1760.720 | (B) patched | 377231.273 | 376369.000 | 1874.971 | | -3.58% | -3.71% | | page_fault3_scalability | | | (A) base | 0.60369 | 0.60072 | 0.0083029 | (B) patched | 0.61733 | 0.61544 | 0.009855 | | +2.26% | +2.45% | | All regressions seem to be minimal, and within the normal variance for the benchmark. The fix for [1] assumes that 3% is noise -- and there were no further practical complaints), so hopefully this means that such variations in these microbenchmarks do not reflect on practical workloads. (3) I also ran stress-ng in a nested cgroup and did not observe any obvious regressions. [1]https://lore.kernel.org/all/20190520063534.GB19312@shao2-debian/
