Re: [PATCH 16/19] sched/numa: Detect if node actively handling migration

[Srikar Dronamraju <srikar@linux.vnet.ibm.com>]

* Rik van Riel <riel@surriel.com> [2018-06-04 16:05:55]:

> On Mon, 2018-06-04 at 15:30 +0530, Srikar Dronamraju wrote:
> 
> > @@ -1554,6 +1562,9 @@ static void task_numa_compare(struct
> > task_numa_env *env,
> >  	if (READ_ONCE(dst_rq->numa_migrate_on))
> >  		return;
> >  
> > +	if (*move && READ_ONCE(pgdat->active_node_migrate))
> > +		*move = false;
> 
> Why not do this check in task_numa_find_cpu?
> 
> That way you won't have to pass this in as a
> pointer, and you also will not have to recalculate
> NODE_DATA(cpu_to_node(env->dst_cpu)) for every CPU.
> 

I thought about this. Lets say we evaluated that destination node can
allow movement. While we iterate through the list of cpus trying to find
the best cpu node, we find a idle cpu towards the end of the list.
However if another task as already raced with us to move a task to this
node, then we should bail out. Keeping the check in task_numa_compare
will allow us to do this.

> >  	/*
> > +	 * If the numa importance is less than SMALLIMP,
> 
>               ^^^ numa improvement
> 

okay

> > +	 * task migration might only result in ping pong
> > +	 * of tasks and also hurt performance due to cache
> > +	 * misses.
> > +	 */
> > +	if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
> > +		goto unlock;
> 
> I can see a use for the first test, but why limit the
> search for the best score once you are past the
> threshold?
> 
> I don't understand the use for that second test.
> 

Lets say few threads are racing with each other to find a cpu on the
node. The first thread has already found a task/cpu 'A' to swap and
finds another task/cpu 'B' thats slightly better than the current
best_cpu which is 'A'. Currently we allow the second task/cpu 'B' to be
set as best_cpu. However the second or subsequent threads cannot find
the first task/cpu A because its suppose to be in active migration. By
the time it reaches task/cpu B even that may look to be in active
migration. It may never know that task/cpu A was cleared. In this way,
the second and subsequent threads may not get a task/cpu in the
preferred node to swap just because the first task kept hopping task/cpu
as its choice of migration.

While we can't complete avoid this, the second check will try to make
sure we don't hop on/hop off just for small incremental numa
improvement.



> What workload benefits from it?

> 
> -- 
> All Rights Reversed.