Re: [RFC PATCH 3/3] sched: Implement shared wakequeue in CFS

[David Vernet <void@manifault.com>]

On Fri, Jun 16, 2023 at 08:53:38AM +0800, Aaron Lu wrote:
> On Thu, Jun 15, 2023 at 06:26:05PM -0500, David Vernet wrote:
>  
> > Ok, it seems that the issue is that I wasn't creating enough netperf
> > clients. I assumed that -n $(nproc) was sufficient. I was able to repro
> 
> Yes that switch is confusing.
> 
> > the contention on my 26 core / 52 thread skylake client as well:
> > 
> > 
>  
> > Thanks for the help in getting the repro on my end.
> 
> You are welcome.
> 
> > So yes, there is certainly a scalability concern to bear in mind for
> > swqueue for LLCs with a lot of cores. If you have a lot of tasks quickly
> > e.g. blocking and waking on futexes in a tight loop, I expect a similar
> > issue would be observed.
> > 
> > On the other hand, the issue did not occur on my 7950X. I also wasn't
> 
> Using netperf/UDP_RR?

Correct

> > able to repro the contention on the Skylake if I ran with the default
> > netperf workload rather than UDP_RR (even with the additional clients).
> 
> I also tried that on the 18cores/36threads/LLC Skylake and the contention
> is indeed much smaller than UDP_RR:
> 
>      7.30%     7.29%  [kernel.vmlinux]      [k]      native_queued_spin_lock_slowpath
> 
> But I wouldn't say it's entirely gone. Also consider Skylake has a lot
> fewer cores per LLC than later Intel servers like Icelake and Sapphire
> Rapids and I expect things would be worse on those two machines.

I cannot reproduce this contention locally, even on a slightly larger
Skylake. Not really sure what to make of the difference here. Perhaps
it's because you're running with CONFIG_SCHED_CORE=y? What is the
change in throughput when you run the default workload on your SKL?

> > I didn't bother to take the mean of all of the throughput results
> > between NO_SWQUEUE and SWQUEUE, but they looked roughly equal.
> > 
> > So swqueue isn't ideal for every configuration, but I'll echo my
> > sentiment from [0] that this shouldn't on its own necessarily preclude
> > it from being merged given that it does help a large class of
> > configurations and workloads, and it's disabled by default.
> > 
> > [0]: https://lore.kernel.org/all/20230615000103.GC2883716@maniforge/
> 
> I was wondering: does it make sense to do some divide on machines with
> big LLCs? Like converting the per-LLC swqueue to per-group swqueue where
> the group can be made of ~8 cpus of the same LLC. This will have a
> similar effect of reducing the number of CPUs in a single LLC so the
> scalability issue can hopefully be fixed while at the same time, it
> might still help some workloads. I realized this isn't ideal in that
> wakeup happens at LLC scale so the group thing may not fit very well
> here.
> 
> Just a thought, feel free to ignore it if you don't think this is
> feasible :-)

That's certainly an idea we could explore, but my inclination would be
to keep everything at a per-LLC granularity. It makes it easier to
reason about performance; both in terms of work conservation per-LLC
(again, not every workload suffers from having large LLCs even if others
do, and halving the size of a swqueue in an LLC could harm other
workloads which benefit from the increased work conservation), and in
terms of contention. To the latter point, I think it would be difficult
to choose an LLC size that wasn't somewhat artificial and workload
specific. If someone has that requirement, I think sched_ext would be a
better alternative.

Thanks,
David