Re: [net] 4890b686f4: netperf.Throughput_Mbps -69.4% regression

[Eric Dumazet]

On Fri, Jun 24, 2022 at 8:01 AM Feng Tang <feng.tang@xxxxxxxxx> wrote:
>
> On Fri, Jun 24, 2022 at 07:45:00AM +0200, Eric Dumazet wrote:
> > On Fri, Jun 24, 2022 at 7:14 AM Feng Tang <feng.tang@xxxxxxxxx> wrote:
> > >
> > > Hi Eric,
> > >
> > > On Fri, Jun 24, 2022 at 06:13:51AM +0200, Eric Dumazet wrote:
> > > > On Fri, Jun 24, 2022 at 3:57 AM Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Thu, 23 Jun 2022 18:50:07 -0400 Xin Long wrote:
> > > > > > From the perf data, we can see __sk_mem_reduce_allocated() is the one
> > > > > > using CPU the most more than before, and mem_cgroup APIs are also
> > > > > > called in this function. It means the mem cgroup must be enabled in
> > > > > > the test env, which may explain why I couldn't reproduce it.
> > > > > >
> > > > > > The Commit 4890b686f4 ("net: keep sk->sk_forward_alloc as small as
> > > > > > possible") uses sk_mem_reclaim(checking reclaimable >= PAGE_SIZE) to
> > > > > > reclaim the memory, which is *more frequent* to call
> > > > > > __sk_mem_reduce_allocated() than before (checking reclaimable >=
> > > > > > SK_RECLAIM_THRESHOLD). It might be cheap when
> > > > > > mem_cgroup_sockets_enabled is false, but I'm not sure if it's still
> > > > > > cheap when mem_cgroup_sockets_enabled is true.
> > > > > >
> > > > > > I think SCTP netperf could trigger this, as the CPU is the bottleneck
> > > > > > for SCTP netperf testing, which is more sensitive to the extra
> > > > > > function calls than TCP.
> > > > > >
> > > > > > Can we re-run this testing without mem cgroup enabled?
> > > > >
> > > > > FWIW I defer to Eric, thanks a lot for double checking the report
> > > > > and digging in!
> > > >
> > > > I did tests with TCP + memcg and noticed a very small additional cost
> > > > in memcg functions,
> > > > because of suboptimal layout:
> > > >
> > > > Extract of an internal Google bug, update from June 9th:
> > > >
> > > > --------------------------------
> > > > I have noticed a minor false sharing to fetch (struct
> > > > mem_cgroup)->css.parent, at offset 0xc0,
> > > > because it shares the cache line containing struct mem_cgroup.memory,
> > > > at offset 0xd0
> > > >
> > > > Ideally, memcg->socket_pressure and memcg->parent should sit in a read
> > > > mostly cache line.
> > > > -----------------------
> > > >
> > > > But nothing that could explain a "-69.4% regression"
> > >
> > > We can double check that.
> > >
> > > > memcg has a very similar strategy of per-cpu reserves, with
> > > > MEMCG_CHARGE_BATCH being 32 pages per cpu.
> > >
> > > We have proposed patch to increase the batch numer for stats
> > > update, which was not accepted as it hurts the accuracy and
> > > the data is used by many tools.
> > >
> > > > It is not clear why SCTP with 10K writes would overflow this reserve constantly.
> > > >
> > > > Presumably memcg experts will have to rework structure alignments to
> > > > make sure they can cope better
> > > > with more charge/uncharge operations, because we are not going back to
> > > > gigantic per-socket reserves,
> > > > this simply does not scale.
> > >
> > > Yes, the memcg statitics and charge/unchage update is very sensitive
> > > with the data alignemnt layout, and can easily trigger peformance
> > > changes, as we've seen quite some similar cases in the past several
> > > years.
> > >
> > > One pattern we've seen is, even if a memcg stats updating or charge
> > > function only takes about 2%~3% of the CPU cycles in perf-profile data,
> > > once it got affected, the peformance change could be amplified to up to
> > > 60% or more.
> > >
> >
> > Reorganizing "struct mem_cgroup" to put "struct page_counter memory"
> > in a separate cache line would be beneficial.
>
> That may help.
>
> And I also want to say the benchmarks(especially micro one) are very
> sensitive to the layout of mem_cgroup. As the 'page_counter' is 112
> bytes in size, I recently made a patch to make it cacheline aligned
> (take 2 cachelines), which improved some hackbench/netperf test
> cases, but caused huge (49%) drop for some vm-scalability tests.
>
> > Many low hanging fruits, assuming nobody will use __randomize_layout on it ;)
> >
> > Also some fields are written even if their value is not changed.
> >
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index abec50f31fe64100f4be5b029c7161b3a6077a74..53d9c1e581e78303ef73942e2b34338567987b74
> > 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -7037,10 +7037,12 @@ bool mem_cgroup_charge_skmem(struct mem_cgroup
> > *memcg, unsigned int nr_pages,
> >                 struct page_counter *fail;
> >
> >                 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
> > -                       memcg->tcpmem_pressure = 0;
> > +                       if (READ_ONCE(memcg->tcpmem_pressure))
> > +                               WRITE_ONCE(memcg->tcpmem_pressure, 0);
> >                         return true;
> >                 }
> > -               memcg->tcpmem_pressure = 1;
> > +               if (!READ_ONCE(memcg->tcpmem_pressure))
> > +                       WRITE_ONCE(memcg->tcpmem_pressure, 1);
> >                 if (gfp_mask & __GFP_NOFAIL) {
> >                         page_counter_charge(&memcg->tcpmem, nr_pages);
> >                         return true;
>
> I will also try this patch, which may take some time.

Note that applications can opt-in reserving memory for one socket,
using SO_RESERVE_MEM

This can be used for jobs with a controlled number of sockets, as this
will avoid many charge/uncharge operations.