Skip to contentSkip to conversation

sched: avoid div in rebalance_tick

PATCH In review Nick Piggin 2 participants 4 messages Jan 12, 2007 – Jan 13, 2007

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sched: avoid div in rebalance_tick

Just noticed this while looking at a bug.

--

Avoid an expensive integer divide 3 times per CPU per tick.

kernel/sched.c +6 -3
Index: linux-2.6/kernel/sched.c
===================================================================
@@ -2887,13 +2887,14 @@ static void active_load_balance(struct r
 static void update_load(struct rq *this_rq)
 {
 	unsigned long this_load;
-	int i, scale;
+	int i;
 
 	this_load = this_rq->raw_weighted_load;
 
 	/* Update our load: */
-	for (i = 0, scale = 1; i < 3; i++, scale <<= 1) {
+	for (i = 0; i < 3; i++) {
 		unsigned long old_load, new_load;
+		int scale;
 
 		old_load = this_rq->cpu_load[i];
 		new_load = this_load;
@@ -2902,9 +2903,11 @@ static void update_load(struct rq *this_
 		 * prevents us from getting stuck on 9 if the load is 10, for
 		 * example.
 		 */
+		scale = 1 << i;
 		if (new_load > old_load)
 			new_load += scale-1;
-		this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
+		this_rq->cpu_load[i] = (old_load*(scale-1) + new_load)
+					>> i; /* (divide by 'scale') */
 	}
 }
 
#1 lore raw
Quoted 2 lines · On Fri, 12 Jan 2007 07:02:13 +0100 Nick Piggin <[email protected]> wrote:

Just noticed this while looking at a bug.
Avoid an expensive integer divide 3 times per CPU per tick.

Integer divide is cheap on some modern processors, and multibit shift isn't on all embedded ones.

How about putting back scale = 1 and using

scale += scale;

instead of the shift and getting what ought to be even better results

#2 lore raw
Quoted 8 lines · Alan
Quoted 2 lines · On Fri, 12 Jan 2007 07:02:13 +0100 Nick Piggin <[email protected]> wrote:

Just noticed this while looking at a bug.
Avoid an expensive integer divide 3 times per CPU per tick.

Integer divide is cheap on some modern processors, and multibit shift isn't on all embedded ones.

Well integer divide unit is non-pipelined on P4 K8 Core2 and probably most processors, AFAIK. So the 3 divs would take 240 cycles on a P4, perhaps.

Quoted 5 lines

How about putting back scale = 1 and using

scale += scale;

instead of the shift and getting what ought to be even better results

Yes I gues we ccan do this as well, good idea. I'll make a quick userspace benchmark and post some numbers with my next submission.

#3 lore raw
In reply to Alan
Quoted 6 lines · Alan
Quoted 2 lines · On Fri, 12 Jan 2007 07:02:13 +0100 Nick Piggin <[email protected]> wrote:

Just noticed this while looking at a bug.
Avoid an expensive integer divide 3 times per CPU per tick.

Quoted 8 lines

Integer divide is cheap on some modern processors, and multibit shift isn't on all embedded ones.

How about putting back scale = 1 and using

scale += scale;

instead of the shift and getting what ought to be even better results

OK, how about this? It only works out to be around 0.01% of my P3's CPU time at 1000HZ, but it also did make the x86 code 16 bytes smaller.

--
Avoid expensive integer divide 3 times per CPU per tick.

A userspace test of this loop went from 26ns, down to 19ns on a G5; and from 123ns down to 28ns on a P3.

(Also avoid a variable bit shift, as suggested by Alan. The effect of this wasn't noticable on the CPUs I tested with).

kernel/sched.c +5 -3
Index: linux-2.6/kernel/sched.c
===================================================================
@@ -2887,14 +2887,16 @@ static void active_load_balance(struct r
 static void update_load(struct rq *this_rq)
 {
 	unsigned long this_load;
-	int i, scale;
+	unsigned int i, scale;
 
 	this_load = this_rq->raw_weighted_load;
 
 	/* Update our load: */
-	for (i = 0, scale = 1; i < 3; i++, scale <<= 1) {
+	for (i = 0, scale = 1; i < 3; i++, scale += scale) {
 		unsigned long old_load, new_load;
 
+		/* scale is effectively 1 << i now, and >> i divides by scale */
+
 		old_load = this_rq->cpu_load[i];
 		new_load = this_load;
 		/*
@@ -2904,7 +2906,7 @@ static void update_load(struct rq *this_
 		 */
 		if (new_load > old_load)
 			new_load += scale-1;
-		this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) / scale;
+		this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
 	}
 }