[lkml]   [2021]   [Feb]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] sched/fair: Prefer idle CPU to cache affinity
On Fri, 2021-02-26 at 22:10 +0530, Srikar Dronamraju wrote:

> Current order of preference to pick a LLC while waking a wake-affine
> task:
> 1. Between the waker CPU and previous CPU, prefer the LLC of the CPU
> that is idle.
> 2. Between the waker CPU and previous CPU, prefer the LLC of the CPU
> that is less lightly loaded.
> In the current situation where waker and previous CPUs are busy, but
> only one of its LLC has an idle CPU, Scheduler may end up picking a
> with no idle CPUs. To mitigate this, add a new step between 1 and 2
> where Scheduler compares idle CPUs in waker and previous LLCs and
> picks
> the appropriate one.

I like that idea a lot. That could also solve some of the
issues sometimes observed on multi-node x86 systems, and
probably on the newer AMD chips with several LLCs on chip.

> + if (sched_feat(WA_WAKER) && tnr_busy < tllc_size)
> + return this_cpu;

I wonder if we need to use a slightly lower threshold on
very large LLCs, both to account for the fact that the
select_idle_cpu code may not find the single idle CPU
among a dozen busy ones, or because on a system with
hyperthreading we may often be better off picking another
LLC for HT contention issues?

Maybe we could use "tnr_busy * 4 <
tllc_size * 3" or
something like that?

That way we will only try to find the last 5 idle
in a 22 CPU LLC if the other LLC also has fewer than 6
idle cores.

That might increase our chances of finding an idle CPU
with SIS_PROP enabled, and might allow WA_WAKER to be
true by default.

> + /* For better wakeup latency, prefer idler LLC to cache
> affinity */
> + diff = tnr_busy * pllc_size - sync - pnr_busy * tllc_size;
> + if (!diff)
> + return nr_cpumask_bits;
> + if (diff < 0)
> + return this_cpu;
> +
> + return prev_cpu;
> +}

All Rights Reversed.
[unhandled content-type:application/pgp-signature]
 \ /
  Last update: 2021-02-27 20:57    [W:0.150 / U:26.964 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site