Re: [PATCH] sched,numa cap pte scanning overhead to 3% of run time

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Nov 05, 2015 at 10:56:29AM -0500, Rik van Riel wrote:
> On 11/05/2015 10:34 AM, Peter Zijlstra wrote:
> > On Wed, Nov 04, 2015 at 01:25:15PM -0500, Rik van Riel wrote:
> >> +++ b/kernel/sched/fair.c
> >> @@ -2155,6 +2155,7 @@ void task_numa_work(struct callback_head *work)
> >>  	unsigned long migrate, next_scan, now = jiffies;
> >>  	struct task_struct *p = current;
> >>  	struct mm_struct *mm = p->mm;
> >> +	u64 runtime = p->se.sum_exec_runtime;
> >>  	struct vm_area_struct *vma;
> >>  	unsigned long start, end;
> >>  	unsigned long nr_pte_updates = 0;
> >> @@ -2277,6 +2278,20 @@ void task_numa_work(struct callback_head *work)
> >>  	else
> >>  		reset_ptenuma_scan(p);
> >>  	up_read(&mm->mmap_sem);
> >> +
> >> +	/*
> >> +	 * There is a fundamental mismatch between the runtime based
> >> +	 * NUMA scanning at the task level, and the wall clock time
> >> +	 * NUMA scanning at the mm level. On a severely overloaded
> >> +	 * system, with very large processes, this mismatch can cause
> >> +	 * the system to spend all of its time in change_prot_numa().
> >> +	 * Limit NUMA PTE scanning to 3% of the task's run time, if
> >> +	 * we spent so much time scanning we got rescheduled.
> >> +	 */
> >> +	if (unlikely(p->se.sum_exec_runtime != runtime)) {
> >> +		u64 diff = p->se.sum_exec_runtime - runtime;
> >> +		p->node_stamp += 32 * diff;
> >> +	}
> > 
> > I don't actually see how this does what it says it does
> 
> If we got rescheduled during the assigning of runtime

Or just had a tick. Even if the whole thing took a fraction of a ms but
we got unlucky and got hit by a tick the sum_exec_runtime would get
updated and not match here.

> Advancing the node_stamp by 32x the amount of time
> the task consumed between entering task_numa_work and
> this point should ensure task_numa_work does not get
> queued again until we have used 32x as much time doing
> something else.

> What am I missing?

The above, issue and the fact that I'm really tired and didn't do 1:32 ~
3%.

So the tick scenario can cause a 32*TICK_NSEC delay even though we spend
much less than TICK_NSEC time scanning, dropping th effective rate much
below the 3%.

Not sure it makes sense to do more accurate accounting, but I suppose we
should mention it somewhere.

> >> @@ -2302,7 +2317,7 @@ void task_tick_numa(struct rq *rq, struct task_struct *curr)
> >>  	now = curr->se.sum_exec_runtime;
> >>  	period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
> >>  
> >> -	if (now - curr->node_stamp > period) {
> >> +	if (now > curr->node_stamp + period) {
> >>  		if (!curr->node_stamp)
> >>  			curr->numa_scan_period = task_scan_min(curr);
> >>  		curr->node_stamp += period;

> I can resend this as a separate patch if you prefer.

Yes, its an unrelated fix.