On Tue, 7 Feb 2023 at 02:03, Xi Wang <xii@xxxxxxxxxx> wrote:
>
> On Mon, Feb 6, 2023 at 1:28 AM Vincent Guittot
> <vincent.guittot@xxxxxxxxxx> wrote:
> >
> > On Fri, 3 Feb 2023 at 19:47, Xi Wang <xii@xxxxxxxxxx> wrote:
> > >
> > > On Fri, Feb 3, 2023 at 1:51 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > > >
> > > > On Tue, Jan 31, 2023 at 05:20:32PM -0800, Xi Wang wrote:
> > > > > After load balancing was split into different scenarios, CPU capacity
> > > > > is ignored for the "migrate_task" case, which means a thread can stay
> > > > > on a softirq heavy cpu for an extended amount of time.
> > > > >
> > > > > By comparing nr_running/capacity instead of just nr_running we can add
> > > > > CPU capacity back into "migrate_task" decisions. This benefits
> > > > > workloads running on machines with heavy network traffic. The change
> > > > > is unlikely to cause serious problems for other workloads but maybe
> > > > > some corner cases still need to be considered.
> > > > >
> > > > > Signed-off-by: Xi Wang <xii@xxxxxxxxxx>
> > > > > ---
> > > > > kernel/sched/fair.c | 3 ++-
> > > > > 1 file changed, 2 insertions(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > > > index 0f8736991427..aad14bc04544 100644
> > > > > --- a/kernel/sched/fair.c
> > > > > +++ b/kernel/sched/fair.c
> > > > > @@ -10368,8 +10368,9 @@ static struct rq *find_busiest_queue(struct lb_env *env,
> > > > > break;
> > > > >
> > > > > case migrate_task:
> > > > > - if (busiest_nr < nr_running) {
> > > > > + if (busiest_nr * capacity < nr_running * busiest_capacity) {
> > > > > busiest_nr = nr_running;
> > > > > + busiest_capacity = capacity;
> > > > > busiest = rq;
> > > > > }
> > > > > break;
> > > >
> > > > I don't think this is correct. The migrate_task case is work-conserving,
> > > > and your change can severely break that I think.
> > > >
> > >
> > > I think you meant this kind of scenario:
> > > cpu 0: idle
> > > cpu 1: 2 tasks
> > > cpu 2: 1 task but only has 30% of capacity
> > > Pulling from cpu 2 is good for the task but lowers the overall cpu
> > > throughput.
> > >
> > > The problem we have is:
> > > cpu 0: idle
> > > cpu 1: 1 task
> > > cpu 2: 1 task but only has 60% of capacity due to net softirq
> > > The task on cpu 2 stays there and runs slower. (This can also be
> > > considered non work-conserving if we account softirq like a task.)
> >
> > When load_balance runs for this 2 cpus, cpu2 should be tagged as
> > misfit_task because of reduce_capacity and should be selected in
> > priority by cpu0 to pull the task. Do you have more details on your
> > topology ?
>
> The topology is 64 core AMD with 2 hyperthreads.
>
> I am not familiar with the related code but I think there are cases
> where a task fits cpu capacity but it can still run faster elsewhere,
> e.g.: Bursty workloads. Thread pool threads with variable utilization
> because it would process more or less requests based on cpu
> availability (pick the next request from a shared queue when the
> previous one is done). A thread having enough cpu cycles but runs
> slower due to softirqs can also directly affect application
> performance.
misfit_task is also used to detect CPUs with reduced capacity. We can
have a look at update_sg_lb_stats(). Your use case above should fall
in this case
>
> > >
> > > Maybe the logic can be merged like this: Use capacity but pick from
> > > nr_running > 1 cpus first, then nr_running == 1 cpus if not found.
