Re: [PATCH v2] sched: push rt tasks only if newly activatedtaskshave been added

From: Gregory Haskins
Date: Mon Apr 21 2008 - 12:34:42 EST


>>> On Mon, Apr 21, 2008 at 12:20 PM, in message <20080421162032.GA12645@xxxxxxx>,
Ingo Molnar <mingo@xxxxxxx> wrote:

> got a build failure:
>
> In file included from kernel/sched.c:1972:
> kernel/sched_rt.c: In function 'enqueue_task_rt':
> kernel/sched_rt.c:522: error: 'struct rt_rq' has no member named 'pushed'
>
> with this config:
>
> http://redhat.com/~mingo/misc/config-Mon_Apr_21_18_02_45_CEST_2008.bad
>
> with the patch below against sched-devel/latest.

Ah, sorry. I had written it against 2.6.25 so I suspect a fuzz issue. I will apply to sched-devel and resolve the errors.

>
> Ingo
>
> ----------->
> Subject: sched: push rt tasks only if newly activated tasks have been added
> From: Gregory Haskins <ghaskins@xxxxxxxxxx>
> Date: Mon, 21 Apr 2008 11:42:18 -0400
>
> SCHED_RR can context-switch many times without having changed the
> run-queue. Therefore trying to push on each context switch can just be
> wasted effort since if it failed the first time, it will likely fail any
> subsequent times as well. Instead, set a flag when we have successfully
> pushed as many tasks away as possible, and only clear it when the
> runqueue adds new tasks (effectively becoming a run-queue "dirty bit").
> If new tasks are added we should try again. If any remote run-queues
> downgrade their priority in the meantime, they will try to pull from us
> (as they always do).
>
> This attempts to address a performance regression reported by Suresh
> Siddha, et. al. in the 2.6.25 series.
>
> Signed-off-by: Gregory Haskins <ghaskins@xxxxxxxxxx>
> CC: suresh.b.siddha@xxxxxxxxx
> CC: mingo@xxxxxxx
> CC: rostedt@xxxxxxxxxxx
> CC: chinang.ma@xxxxxxxxx
> CC: arjan@xxxxxxxxxxxxxxx
> CC: willy@xxxxxxxxxxxxxxx
> Signed-off-by: Ingo Molnar <mingo@xxxxxxx>
> ---
> kernel/sched.c | 2 ++
> kernel/sched_rt.c | 22 ++++++++++++++++++++--
> 2 files changed, 22 insertions(+), 2 deletions(-)
>
> Index: linux/kernel/sched.c
> ===================================================================
> --- linux.orig/kernel/sched.c
> +++ linux/kernel/sched.c
> @@ -462,6 +462,7 @@ struct rt_rq {
> #ifdef CONFIG_SMP
> unsigned long rt_nr_migratory;
> int overloaded;
> + int pushed;
> #endif
> int rt_throttled;
> u64 rt_time;
> @@ -8073,6 +8074,7 @@ static void init_rt_rq(struct rt_rq *rt_
> #ifdef CONFIG_SMP
> rt_rq->rt_nr_migratory = 0;
> rt_rq->overloaded = 0;
> + rt_rq->pushed = 0;
> #endif
>
> rt_rq->rt_time = 0;
> Index: linux/kernel/sched_rt.c
> ===================================================================
> --- linux.orig/kernel/sched_rt.c
> +++ linux/kernel/sched_rt.c
> @@ -514,6 +514,13 @@ static void enqueue_task_rt(struct rq *r
> for_each_sched_rt_entity(rt_se)
> enqueue_rt_entity(rt_se);
>
> + /*
> + * Clear the "pushed" state since we have changed the run-queue and
> + * may need to migrate some of these tasks (see push_rt_tasks() for
> + * details)
> + */
> + rq->rt.pushed = 0;
> +
> inc_cpu_load(rq, p->se.load.weight);
> }
>
> @@ -913,7 +920,7 @@ static int push_rt_task(struct rq *rq)
> int ret = 0;
> int paranoid = RT_MAX_TRIES;
>
> - if (!rq->rt.overloaded)
> + if (!rq->rt.overloaded || rq->rt.pushed)
> return 0;
>
> next_task = pick_next_highest_task_rt(rq, -1);
> @@ -973,6 +980,15 @@ out:
> }
>
> /*
> + * push_rt_tasks()
> + *
> + * Push as many tasks away as possible. We only need to try once whenever
> + * one or more new tasks are added to our runqueue. After an inital
> attempt,
> + * further changes in remote runqueue state will be accounted for with pull
> + * operations. Therefore, we mark the run-queue as "pushed" here, and clear
> it
> + * during enqueue. This primarily helps SCHED_RR tasks which will tend to
> + * have a higher context-switch to enqueue ratio.
> + *
> * TODO: Currently we just use the second highest prio task on
> * the queue, and stop when it can't migrate (or there's
> * no more RT tasks). There may be a case where a lower
> @@ -987,6 +1003,8 @@ static void push_rt_tasks(struct rq *rq)
> /* push_rt_task will return true if it moved an RT */
> while (push_rt_task(rq))
> ;
> +
> + rq->rt.pushed = 1;
> }
>
> static int pull_rt_task(struct rq *this_rq)
> @@ -1091,7 +1109,7 @@ static void post_schedule_rt(struct rq *
> * the lock was owned by prev, we need to release it
> * first via finish_lock_switch and then reaquire it here.
> */
> - if (unlikely(rq->rt.overloaded)) {
> + if (unlikely(rq->rt.overloaded && !rq->rt.pushed)) {
> spin_lock_irq(&rq->lock);
> push_rt_tasks(rq);
> spin_unlock_irq(&rq->lock);
> --
> To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/