On Thu, May 11, 2023 at 9:00 AM Zhang, Cathy <cathy.zhang@xxxxxxxxx> wrote:
>
>
>
> > -----Original Message-----
> > From: Zhang, Cathy
> > Sent: Thursday, May 11, 2023 8:53 AM
> > To: Shakeel Butt <shakeelb@xxxxxxxxxx>
> > Cc: Eric Dumazet <edumazet@xxxxxxxxxx>; Linux MM <linux-
> > mm@xxxxxxxxx>; Cgroups <cgroups@xxxxxxxxxxxxxxx>; Paolo Abeni
> > <pabeni@xxxxxxxxxx>; davem@xxxxxxxxxxxxx; kuba@xxxxxxxxxx;
> > Brandeburg, Jesse <jesse.brandeburg@xxxxxxxxx>; Srinivas, Suresh
> > <suresh.srinivas@xxxxxxxxx>; Chen, Tim C <tim.c.chen@xxxxxxxxx>; You,
> > Lizhen <Lizhen.You@xxxxxxxxx>; eric.dumazet@xxxxxxxxx;
> > netdev@xxxxxxxxxxxxxxx
> > Subject: RE: [PATCH net-next 1/2] net: Keep sk->sk_forward_alloc as a proper
> > size
> >
> >
> >
> > > -----Original Message-----
> > > From: Shakeel Butt <shakeelb@xxxxxxxxxx>
> > > Sent: Thursday, May 11, 2023 3:00 AM
> > > To: Zhang, Cathy <cathy.zhang@xxxxxxxxx>
> > > Cc: Eric Dumazet <edumazet@xxxxxxxxxx>; Linux MM <linux-
> > > mm@xxxxxxxxx>; Cgroups <cgroups@xxxxxxxxxxxxxxx>; Paolo Abeni
> > > <pabeni@xxxxxxxxxx>; davem@xxxxxxxxxxxxx; kuba@xxxxxxxxxx;
> > Brandeburg,
> > > Jesse <jesse.brandeburg@xxxxxxxxx>; Srinivas, Suresh
> > > <suresh.srinivas@xxxxxxxxx>; Chen, Tim C <tim.c.chen@xxxxxxxxx>; You,
> > > Lizhen <lizhen.you@xxxxxxxxx>; eric.dumazet@xxxxxxxxx;
> > > netdev@xxxxxxxxxxxxxxx
> > > Subject: Re: [PATCH net-next 1/2] net: Keep sk->sk_forward_alloc as a
> > > proper size
> > >
> > > On Wed, May 10, 2023 at 9:09 AM Zhang, Cathy <cathy.zhang@xxxxxxxxx>
> > > wrote:
> > > >
> > > >
> > > [...]
> > > > > > >
> > > > > > > Have you tried to increase batch sizes ?
> > > > > >
> > > > > > I jus picked up 256 and 1024 for a try, but no help, the
> > > > > > overhead still
> > > exists.
> > > > >
> > > > > This makes no sense at all.
> > > >
> > > > Eric,
> > > >
> > > > I added a pr_info in try_charge_memcg() to print nr_pages if
> > > > nr_pages
> > > > >= MEMCG_CHARGE_BATCH, except it prints 64 during the initialization
> > > > of instances, there is no other output during the running. That
> > > > means nr_pages is not over 64, I guess that might be the reason why
> > > > to increase MEMCG_CHARGE_BATCH doesn't affect this case.
> > > >
> > >
> > > I am assuming you increased MEMCG_CHARGE_BATCH to 256 and 1024
> > but
> > > that did not help. To me that just means there is a different
> > > bottleneck in the memcg charging codepath. Can you please share the
> > > perf profile? Please note that memcg charging does a lot of other
> > > things as well like updating memcg stats and checking (and enforcing)
> > > memory.high even if you have not set memory.high.
> >
> > Thanks Shakeel! I will check more details on what you mentioned. We use
> > "sudo perf top -p $(docker inspect -f '{{.State.Pid}}' memcached_2)" to
> > monitor one of those instances, and also use "sudo perf top" to check the
> > overhead from system wide.
>
> Here is the annotate output of perf top for the three memcg hot paths:
>
> Showing cycles for page_counter_try_charge
> Events Pcnt (>=5%)
> Percent | Source code & Disassembly of elf for cycles (543288 samples, percent: local period)
> ---------------------------------------------------------------------------------------------------
> 0.00 : ffffffff8141388d: mov %r12,%rax
> 76.82 : ffffffff81413890: lock xadd %rax,(%rbx)
> 22.10 : ffffffff81413895: lea (%r12,%rax,1),%r15
>
>
> Showing cycles for page_counter_cancel
> Events Pcnt (>=5%)
> Percent | Source code & Disassembly of elf for cycles (1004744 samples, percent: local period)
> ----------------------------------------------------------------------------------------------------
> : 160 return i + xadd(&v->counter, i);
> 77.42 : ffffffff81413759: lock xadd %rax,(%rdi)
> 22.34 : ffffffff8141375e: sub %rsi,%rax
>
>
> Showing cycles for try_charge_memcg
> Events Pcnt (>=5%)
> Percent | Source code & Disassembly of elf for cycles (256531 samples, percent: local period)
> ---------------------------------------------------------------------------------------------------
> : 22 return __READ_ONCE((v)->counter);
> 77.53 : ffffffff8141df86: mov 0x100(%r13),%rdx
> : 2826 READ_ONCE(memcg->memory.high);
> 19.45 : ffffffff8141df8d: mov 0x190(%r13),%rcx
This is rephrasing the info you gave earlier ?
16.77% [kernel] [k] page_counter_try_charge
16.56% [kernel] [k] page_counter_cancel
15.65% [kernel] [k] try_charge_memcg
What matters here is a call graph.
perf record -a -g sleep 5 # While the test is running
perf report --no-children --stdio
What precise kernel are you using btw ?
