[patch][rfc] quell interactive feeding frenzy

[Mike Galbraith <efault@gmx.de>]

Greetings,

Problem:  Wake-up -> cpu latency increases with the number of runnable
tasks, ergo adding this latency to sleep_avg becomes increasingly potent
as nr_running increases.  This turns into a very nasty problem with as
few as 10 httpd tasks doing round robin scheduling.  The result is that
you can only login with difficulty, and interactivity is nonexistent.

Solution:  Restrict the amount of boost a task can receive from this
mechanism, and disable the mechanism entirely when load is high.  As
always, there is a price for increasing fairness.  In this case, the
price seems worth it.  It bought me a usable 2.6 apache server.


Signed-off-by: Mike Galbraith <efault@gmx.de>

--- linux-2.6.17-rc1/kernel/sched.c.org	2006-04-07 08:52:47.000000000 +0200
+++ linux-2.6.17-rc1/kernel/sched.c	2006-04-07 08:57:34.000000000 +0200
@@ -3012,14 +3012,20 @@ go_idle:
 	queue = array->queue + idx;
 	next = list_entry(queue->next, task_t, run_list);
 
-	if (!rt_task(next) && interactive_sleep(next->sleep_type)) {
+	if (!rt_task(next) && interactive_sleep(next->sleep_type) &&
+			rq->nr_running < 1 + MAX_BONUS - CURRENT_BONUS(next)) {
 		unsigned long long delta = now - next->timestamp;
+		unsigned long max_delta;
 		if (unlikely((long long)(now - next->timestamp) < 0))
 			delta = 0;
 
 		if (next->sleep_type == SLEEP_INTERACTIVE)
 			delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
 
+		max_delta = (1 + MAX_BONUS - CURRENT_BONUS(next)) * GRANULARITY;
+		max_delta = JIFFIES_TO_NS(max_delta);
+		if (delta > max_delta)
+			delta = max_delta;
 		array = next->array;
 		new_prio = recalc_task_prio(next, next->timestamp + delta);
 


-
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/