Re: [git] CFS-devel, group scheduler, fixes

[Tong Li <tong.n.li@intel.com>]

Mike,

Could you try this patch to see if it solves the latency problem?

  tong

Changes:

1. Modified vruntime adjustment logic in set_task_cpu(). See comments in 
code. This fixed the negative vruntime problem.

2. This code in update_curr() seems to be wrong:

if (unlikely(!curr))
	return sync_vruntime(cfs_rq);

cfs_rq->curr can be NULL even if cfs_rq->nr_running is non-zero (e.g., 
when an RT task is running). We only want to call sync_vruntime when 
cfs_rq->nr_running is 0. This fixed the large latency problem (at least in 
my tests).

Signed-off-by: Tong Li <tong.n.li@intel.com>
---
diff -uprN linux-2.6-sched-devel-orig/kernel/sched.c linux-2.6-sched-devel/kernel/sched.c
--- linux-2.6-sched-devel-orig/kernel/sched.c	2007-09-20 12:15:41.000000000 -0700
+++ linux-2.6-sched-devel/kernel/sched.c	2007-09-21 19:40:08.000000000 -0700
@@ -1033,9 +1033,20 @@ void set_task_cpu(struct task_struct *p,
 	if (p->se.block_start)
 		p->se.block_start -= clock_offset;
 #endif
-	if (likely(new_rq->cfs.min_vruntime))
-		p->se.vruntime -= old_rq->cfs.min_vruntime -
-						new_rq->cfs.min_vruntime;
+	/* 
+	 * Reset p's vruntime if it moves to new_cpu whose min_vruntime is 
+	 * 100,000,000 (equivalent to 100ms for nice-0 tasks) larger or
+	 * smaller than p's vruntime. This improves interactivity when
+	 * pinned and unpinned tasks co-exist. For example, pinning a few
+	 * tasks to a CPU can cause its min_vruntime much smaller than the
+	 * other CPUs. If a task moves to this CPU, its vruntime can be so
+	 * large it won't be scheduled until the locally pinned tasks'
+	 * vruntimes catch up, causing large delays.
+	 */
+	if (unlikely(old_cpu != new_cpu && p->se.vruntime && 
+		(p->se.vruntime > new_rq->cfs.min_vruntime + 100000000 || 
+		 p->se.vruntime + 100000000 < new_rq->cfs.min_vruntime)))
+		p->se.vruntime = new_rq->cfs.min_vruntime;

 	__set_task_cpu(p, new_cpu);
 }
@@ -1599,6 +1610,7 @@ static void __sched_fork(struct task_str
 	p->se.exec_start		= 0;
 	p->se.sum_exec_runtime		= 0;
 	p->se.prev_sum_exec_runtime	= 0;
+	p->se.vruntime			= 0;

 #ifdef CONFIG_SCHEDSTATS
 	p->se.wait_start		= 0;
diff -uprN linux-2.6-sched-devel-orig/kernel/sched_fair.c linux-2.6-sched-devel/kernel/sched_fair.c
--- linux-2.6-sched-devel-orig/kernel/sched_fair.c	2007-09-20 12:15:41.000000000 -0700
+++ linux-2.6-sched-devel/kernel/sched_fair.c	2007-09-21 17:23:09.000000000 -0700
@@ -306,8 +306,10 @@ static void update_curr(struct cfs_rq *c
 	u64 now = rq_of(cfs_rq)->clock;
 	unsigned long delta_exec;

-	if (unlikely(!curr))
+	if (unlikely(!cfs_rq->nr_running))
 		return sync_vruntime(cfs_rq);
+	if (unlikely(!curr))
+		return;

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