Re: [RFC PATCH 1/1] sched/fair: ratelimit update to tg->load_avg

[Aaron Lu]

Hi Gautham,

Thanks for sharing this information.

On Wed, Aug 16, 2023 at 10:12:52AM +0530, Gautham R. Shenoy wrote:
> Hello Aaron,
> (Adding David Vernet)
> 
> On Wed, Aug 16, 2023 at 10:48:31AM +0800, Aaron Lu wrote:
> > When using sysbench to benchmark Postgres in a single docker instance
> > with sysbench's nr_threads set to nr_cpu, it is observed there are times
> > update_cfs_group() and update_load_avg() shows noticeable overhead on
> > a 2sockets/112core/224cpu Intel Sapphire Rapids(SPR):
> > 
> >     13.75%    13.74%  [kernel.vmlinux]           [k] update_cfs_group
> >     10.63%    10.04%  [kernel.vmlinux]           [k] update_load_avg
> > 
> > Annotate shows the cycles are mostly spent on accessing tg->load_avg
> > with update_load_avg() being the write side and update_cfs_group() being
> > the read side. tg->load_avg is per task group and when different tasks
> > of the same taskgroup running on different CPUs frequently access
> > tg->load_avg, it can be heavily contended.
> 
> 
> Interestingly I observed this contention on 2 socket EPYC servers
> (Zen3 and Zen4) while running tbench and netperf with David Vernet's
> shared-runqueue v3 patches. This contention was observed only when
> running with the shared-runqueue enabled but not otherwise.

I think without shared-runqueue, the migration number is pretty low on
EPYC since it has much smaller LLC than Intel's.

> 
>   Overhead  Command  Shared Object     Symbol
> +   20.54%  tbench   [kernel.vmlinux]  [k] update_cfs_group
> +   15.78%  tbench   [kernel.vmlinux]  [k] update_load_avg
> 
> This was causing the tbench (and netperf) to not scale beyond 32
> clients when shared-runqueue was enabled.
> 
> > 
> > The frequent access to tg->load_avg is due to task migration on wakeup
> > path, e.g. when running postgres_sysbench on a 2sockets/112cores/224cpus
> > Intel Sappire Rapids, during a 5s window, the wakeup number is 14millions
> > and migration number is 11millions and with each migration, the task's
> > load will transfer from src cfs_rq to target cfs_rq and each change
> > involves an update to tg->load_avg.
> 
> With the shared-runqueue patches, we see a lot more task migrations
> since the newidle_balance() path would pull tasks from the
> shared-runqueue. While the read of tg->load_avg is via READ_ONCE on
> x86, the write is atomic.

That makes sense. More migrations, more overhead from update_load_avg()
and update_cfs_group().

> 
> > Since the workload can trigger as many
> > wakeups and migrations, the access(both read and write) to tg->load_avg
> > can be unbound. As a result, the two mentioned functions showed noticeable
> > overhead. With netperf/nr_client=nr_cpu/UDP_RR, the problem is worse:
> > during a 5s window, wakeup number is 21millions and migration number is
> > 14millions; update_cfs_group() costs ~25% and update_load_avg() costs ~16%.
> > 
> > Reduce the overhead by limiting updates to tg->load_avg to at most once
> > per ms. After this change, the cost of accessing tg->load_avg is greatly
> > reduced and performance improved. Detailed test results below.
> 
> I will try this patch on with David's series today.

Look forward to see your numbers.

Thanks,
Aaron