Message-ID: <20170803131537.GB17196@destiny>
Date:   Thu, 3 Aug 2017 09:15:38 -0400
From:   Josef Bacik <josef@...icpanda.com>
To:     Brendan Jackman <brendan.jackman@....com>
Cc:     Josef Bacik <josef@...icpanda.com>,
        Mike Galbraith <umgwanakikbuti@...il.com>,
        Joel Fernandes <joelaf@...gle.com>,
        Peter Zijlstra <peterz@...radead.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Juri Lelli <Juri.Lelli@....com>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Patrick Bellasi <patrick.bellasi@....com>,
        Chris Redpath <Chris.Redpath@....com>
Subject: Re: wake_wide mechanism clarification

On Thu, Aug 03, 2017 at 11:53:19AM +0100, Brendan Jackman wrote:
> 
> Hi,
> 
> On Fri, Jun 30 2017 at 17:55, Josef Bacik wrote:
> > On Fri, Jun 30, 2017 at 07:02:20PM +0200, Mike Galbraith wrote:
> >> On Fri, 2017-06-30 at 10:28 -0400, Josef Bacik wrote:
> >> > On Thu, Jun 29, 2017 at 08:04:59PM -0700, Joel Fernandes wrote:
> >> >
> >> > > That makes sense that we multiply slave's flips by a factor because
> >> > > its low, but I still didn't get why the factor is chosen to be
> >> > > llc_size instead of something else for the multiplication with slave
> >> > > (slave * factor).
> >>
> >> > Yeah I don't know why llc_size was chosen...
> >>
> >> static void update_top_cache_domain(int cpu)
> >> {
> >> struct sched_domain_shared *sds = NULL;
> >> struct sched_domain *sd;
> >> int id = cpu;
> >> int size = 1;
> >>
> >> sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
> >> if (sd) {
> >> id = cpumask_first(sched_domain_span(sd));
> >> size = cpumask_weight(sched_domain_span(sd));
> >> sds = sd->shared;
> >> }
> >>
> >> rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
> >> per_cpu(sd_llc_size, cpu) = size;
> >>
> >> The goal of wake wide was to approximate when pulling would be a futile
> >> consolidation effort and counterproductive to scaling. 'course with
> >> ever increasing socket size, any 1:N waker is ever more likely to run
> >> out of CPU for its one and only self (slamming into scaling wall)
> >> before it needing to turn its minions loose to conquer the world.
> >>
> >> Something else to consider: network interrupt waking multiple workers
> >> at high frequency. If the waking CPU is idle, do you really want to
> >> place a worker directly in front of a tattoo artist, or is it better
> >> off nearly anywhere but there?
> >>
> >> If the box is virtual, with no topology exposed (or real but ancient)
> >> to let select_idle_sibling() come to the rescue, two workers can even
> >> get tattooed simultaneously (see sync wakeup).
> >>
> >
> > Heuristics are hard, news at 11.  I think messing with wake_wide() itself is too
> > big of a hammer, we probably need a middle ground.  I'm messing with it right
> > now so it's too early to say for sure, but i _suspect_ the bigger latencies we
> > see are not because we overload the cpu we're trying to pull to, but because
> > when we fail to do the wake_affine() we only look at siblings of the affine_sd
> > instead of doing the full "find the idlest cpu in the land!" thing.
> 
> This is the problem I've been hitting lately. My use case is 1 task per
> CPU on ARM big.LITTLE (asymmetrical CPU capacity). The workload is 1
> task per CPU, they all do X amount of work then pthread_barrier_wait
> (i.e. sleep until the last task finishes its X and hits the barrier). On
> big.LITTLE, the tasks which get a "big" CPU finish faster, and then
> those CPUs pull over the tasks that are still running:
> 
>      v CPU v           ->time->
> 
>                     -------------
>    0  (big)         11111  /333
>                     -------------
>    1  (big)         22222   /444|
>                     -------------
>    2  (LITTLE)      333333/
>                     -------------
>    3  (LITTLE)      444444/
>                     -------------
> 
> Now when task 4 hits the barrier (at |) and wakes the others up, there
> are 4 tasks with prev_cpu=<big> and 0 tasks with
> prev_cpu=<little>. Assuming that those wakeups happen on CPU4,
> regardless of wake_affine, want_affine means that we'll only look in
> sd_llc (cpus 0 and 1), so tasks will be unnecessarily coscheduled on the
> bigs until the next load balance, something like this:
> 
>      v CPU v           ->time->
> 
>                     ------------------------
>    0  (big)         11111  /333  31313\33333
>                     ------------------------
>    1  (big)         22222   /444|424\4444444
>                     ------------------------
>    2  (LITTLE)      333333/          \222222
>                     ------------------------
>    3  (LITTLE)      444444/            \1111
>                     ------------------------
>                                  ^^^
>                            underutilization
> 
> > I _think_
> > the answer is to make select_idle_sibling() try less hard to find something
> > workable and only use obviously idle cpu's in the affine sd, and fall back to
> > the full load balance esque search.
> 
> So this idea of allowing select_idle_sibling to fail, and falling back
> to the slow path, would help me too, I think.

Unfortunately this statement of mine was wrong, I had it in my head that we
would fall back to a find the idlest cpu thing provided we failed to wake
affine, but we just do select_idle_sibling() and expect the load balancer to
move things around as needed.

> 
> This is also why I was playing with your
> don't-affine-recently-balanced-tasks patch[1], which also helps my case
> since it prevents want_affine for tasks 3 and 4 (which were recently
> moved by an active balance).
> 
> [1] https://marc.info/?l=linux-kernel&m=150003849602535&w=2
>     (also linked elsewhere in this thread)
> 

Would you try peter's sched/experimental branch and see how that affects your
workload?  I'm still messing with my patches and I may drop this one as it now
appears to be too aggressive with the new set of patches.  Thanks,

Josef

Powered by blists - more mailing lists