Le lundi 26 oct. 2020 à 08:45:27 (-0400), Chris Mason a écrit :
> On 26 Oct 2020, at 4:39, Vincent Guittot wrote:
>
> > Hi Chris
> >
> > On Sat, 24 Oct 2020 at 01:49, Chris Mason <c...@fb.com> wrote:
> > >
> > > Hi everyone,
> > >
> > > We’re validating a new kernel in the fleet, and compared with v5.2,
> >
> > Which version are you using ?
> > several improvements have been added since v5.5 and the rework of
> > load_balance
>
> We’re validating v5.6, but all of the numbers referenced in this patch are
> against v5.9. I usually try to back port my way to victory on this kind of
> thing, but mainline seems to behave exactly the same as 0b0695f2b34a wrt
> this benchmark.
ok. Thanks for the confirmation
I have been able to reproduce the problem on my setup.
Could you try the fix below ?
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -9049,7 +9049,8 @@ static inline void calculate_imbalance(struct lb_env
*env, struct sd_lb_stats *s
* emptying busiest.
*/
if (local->group_type == group_has_spare) {
- if (busiest->group_type > group_fully_busy) {
+ if ((busiest->group_type > group_fully_busy) &&
+ (busiest->group_weight > 1)) {
/*
* If busiest is overloaded, try to fill spare
* capacity. This might end up creating spare capacity
When we calculate an imbalance at te smallest level, ie between CPUs
(group_weight == 1),
we should try to spread tasks on cpus instead of trying to fill spare capacity.
>
> >
> > > performance is ~2-3% lower for some of our workloads. After some
> > > digging, Johannes found that our involuntary context switch rate was
> > > ~2x
> > > higher, and we were leaving a CPU idle a higher percentage of the
> > > time,
> > > even though the workload was trying to saturate the system.
> > >
> > > We were able to reproduce the problem with schbench, and Johannes
> > > bisected down to:
> > >
> > > commit 0b0695f2b34a4afa3f6e9aa1ff0e5336d8dad912
> > > Author: Vincent Guittot <vincent.guit...@linaro.org>
> > > Date: Fri Oct 18 15:26:31 2019 +0200
> > >
> > > sched/fair: Rework load_balance()
> > >
> > > Our working theory is the load balancing changes are leaving
> > > processes
> > > behind busy CPUs instead of moving them onto idle ones. I made a few
> > > schbench modifications to make this easier to demonstrate:
> > >
> > > https://git.kernel.org/pub/scm/linux/kernel/git/mason/schbench.git/
> > >
> > > My VM has 40 cpus (20 cores, 2 threads per core), and my schbench
> > > command line is:
> >
> > What is the topology ? are they all part of the same LLC ?
>
> We’ve seen the regression on both single socket and dual socket bare metal
> intel systems. On the VM I reproduced with, I saw similar latencies with
> and without siblings configured into the topology.
>
> -chris
