Sacrifice throughput to ensure low latency at all load levels.

Probably not a good thing to do by default.

-ck

--- linux-2.6.35-ck1.orig/kernel/sched_bfs.c	2010-09-23 08:56:29.871350304 +1000
+++ linux-2.6.35-ck1/kernel/sched_bfs.c	2010-09-24 17:08:30.935726514 +1000
@@ -2069,6 +2041,7 @@ update_cpu_clock(struct rq *rq, struct t
 	/* time_slice accounting is done in usecs to avoid overflow on 32bit */
 	if (rq->rq_policy != SCHED_FIFO && p != idle) {
 		long time_diff = rq->clock - rq->rq_last_ran;
+		int qnr = queued_notrunning();

 		/*
 		 * There should be less than or equal to one jiffy worth, and not
@@ -2076,11 +2049,23 @@ update_cpu_clock(struct rq *rq, struct t
 		 * time_slice accounting.
 		 */
 		if (unlikely(time_diff <= 0))
-			time_diff = JIFFIES_TO_NS(1) / 2;
+			time_diff = HALF_JIFFY_NS;
 		else if (unlikely(time_diff > JIFFIES_TO_NS(1)))
 			time_diff = JIFFIES_TO_NS(1);

-		rq->rq_time_slice -= NS_TO_US(time_diff);
+		/*
+		 * If we are overloaded, then shorten the effective timeslices
+		 * to ensure latencies are kept as small as is possible by
+		 * making them expire at a rate proportional to load/CPUs.
+		 */
+		if (qnr > 0) {
+			int nol = num_online_cpus();
+
+			time_diff /= nol;
+			time_diff *= qnr + nol;
+		}
+		time_diff = NS_TO_US(time_diff);
+		rq->rq_time_slice -= time_diff;
 	}
 	rq->rq_last_ran = rq->timekeep_clock = rq->clock;
 }