lkml.org 
[lkml]   [2020]   [Dec]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC PATCH v7] sched/fair: select idle cpu from idle cpumask for task wakeup
On Mon, Dec 14, 2020 at 10:18:16AM +0100, Vincent Guittot wrote:
> On Fri, 11 Dec 2020 at 18:45, Peter Zijlstra <peterz@infradead.org> wrote:
> >
> > On Thu, Dec 10, 2020 at 12:58:33PM +0000, Mel Gorman wrote:
> > > The prequisite patch to make that approach work was rejected though
> > > as on its own, it's not very helpful and Vincent didn't like that the
> > > load_balance_mask was abused to make it effective.
> >
> > So last time I poked at all this, I found that using more masks was a
> > performance issue as well due to added cache misses.
> >
> > Anyway, I finally found some time to look at all this, and while reading
> > over the whole SiS crud to refresh the brain came up with the below...
> >
> > It's still naf, but at least it gets rid of a bunch of duplicate
> > scanning and LoC decreases, so it should be awesome. Ofcourse, as
> > always, benchmarks will totally ruin everything, we'll see, I started
> > some.
> >
> > It goes on top of Mel's first two patches (which is about as far as I
> > got)...
>
> We have several different things that Aubrey and Mel patches are
> trying to achieve:
>
> Aubrey wants to avoid scanning busy cpus
> - patch works well on busy system: I see a significant benefit with
> hackbench on a lot of group on my 2 nodes * 28 cores * 4 smt
> hackbench -l 2000 -g 128
> tip 3.334
> w/ patch 2.978 (+12%)
>

It's still the case that Aubrey's work does not conflict with what Peter
is doing. All that changes is what mask is applied.

Similarly, the p->recent_used_cpu changes I made initially (but got
rejected) reduces scanning. I intend to revisit that to use recent_used_cpu
as a search target because one side-effect of the patch was the siblings
can be used prematurely.

> - Aubey also tried to not scan the cpus which are idle for a short
> duration (when a tick not stopped) but there are problems because not
> stopping a tick doesn't mean a short idle. In fact , something similar
> to find_idlest_group_cpu() should be done in this case but then it's
> no more a fast path search
>

The crowd most likely to complain about this is Facebook as they have
workloads that prefer deep searches to find idle CPUs.

> Mel wants to minimize looping over the cpus
> -patch 4 is an obvious win on light loaded system because it avoids
> looping twice the cpus mask when some cpus are idle but no core

Peter's patch replaces that entirely which I'm fine with.

> -But patch 3 generates perf régression
> hackbench -l 2000 -g 1
> tip 12.293
> /w all Mel's patches 15.163 -14%
> /w Aubrey + Mel patches minus patch 3 : 12.788 +3.8% But I think that
> Aubreay's patch doesn't help here. Test without aubrey's patch are
> running
>
> -he also tried to used load_balance_mask to do something similar to the below
>

Peter's approach removes load_balance_mask abuse so I think it's a
better approach overall to scanning LLC domains in a single pass. It
needs refinement and a lot of testing but I think it's promising.

> > -static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
> > +static int __select_idle_core(int core, struct cpumask *cpus, int *idle_cpu)
> > {
> > - struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
> > - int core, cpu;
> > -
> > - if (!static_branch_likely(&sched_smt_present))
> > - return -1;
> > -
> > - if (!test_idle_cores(target, false))
> > - return -1;
> > -
> > - cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
> > -
> > - for_each_cpu_wrap(core, cpus, target) {
> > - bool idle = true;
> > + bool idle = true;
> > + int cpu;
> >
> > - for_each_cpu(cpu, cpu_smt_mask(core)) {
> > - if (!available_idle_cpu(cpu)) {
> > - idle = false;
> > - break;
> > - }
> > + for_each_cpu(cpu, cpu_smt_mask(core)) {
> > + if (!available_idle_cpu(cpu)) {
> > + idle = false;
> > + continue;
>
> By not breaking on the first not idle cpu of the core, you will
> largely increase the number of loops. On my system, it increases 4
> times from 28 up tu 112
>

But breaking early will mean that SMT siblings are used prematurely.
However, if SIS_PROP was partially enforced for the idle core scan, it
could limit the search for an idle core once an idle candidate was
discovered.

--
Mel Gorman
SUSE Labs

\
 
 \ /
  Last update: 2020-12-14 13:44    [W:0.102 / U:0.056 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site