[PATCH] Speedup divides by cpu_power in scheduler

[Eric Dumazet]

I noticed expensive divides done in try_to_wakeup() and find_busiest_group() 
on a bi dual core Opteron machine (total of 4 cores), moderatly loaded (15.000 
context switch per second)

oprofile numbers :

CPU: AMD64 processors, speed 2600.05 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Cycles outside of halt state) with a unit 
mask of 0x00 (No unit mask) count 50000
samples  %        symbol name
...
613914    1.0498  try_to_wake_up
   834  0.0013 :ffffffff80227ae1:   div    %rcx
77513  0.1191 :ffffffff80227ae4:   mov    %rax,%r11

608893    1.0413  find_busiest_group
  1841  0.0031 :ffffffff802260bf:       div    %rdi
140109  0.2394 :ffffffff802260c2:       test   %sil,%sil


Some of these divides can use the reciprocal divides we introduced some time 
ago (currently used in slab AFAIK)

We can assume a load will fit in a 32bits number, because with a 
SCHED_LOAD_SCALE=128 value, its still a theorical limit of 33554432

When/if we reach this limit one day, probably cpus will have a fast hardware 
divide and we can zap the reciprocal divide trick.

I did not convert the divide in cpu_avg_load_per_task(), because tracking 
nr_running changes may be not worth it ? We could use a static table of 32 
reciprocal values but it would add a conditional branch and table lookup.

Signed-off-by: Eric Dumazet <dada1@xxxxxxxxxxxxx>
--- linux-2.6.21-rc1/include/linux/sched.h	2007-02-21 21:08:32.000000000 +0100
+++ linux-2.6.21-rc1-ed/include/linux/sched.h	2007-02-22 08:53:26.000000000 +0100
@@ -669,7 +669,12 @@ struct sched_group {
 	 * CPU power of this group, SCHED_LOAD_SCALE being max power for a
 	 * single CPU. This is read only (except for setup, hotplug CPU).
 	 */
-	unsigned long cpu_power;
+	unsigned int cpu_power;
+	/*
+	 * reciprocal value of cpu_power to avoid expensive divides
+	 * (see include/linux/reciprocal_div.h)
+	 */
+	u32 reciprocal_cpu_power;
 };
 
 struct sched_domain {
--- linux-2.6.21-rc1/kernel/sched.c.orig	2007-02-21 21:10:54.000000000 +0100
+++ linux-2.6.21-rc1-ed/kernel/sched.c	2007-02-22 08:46:56.000000000 +0100
@@ -52,6 +52,7 @@
 #include <linux/tsacct_kern.h>
 #include <linux/kprobes.h>
 #include <linux/delayacct.h>
+#include <linux/reciprocal_div.h>
 #include <asm/tlb.h>
 
 #include <asm/unistd.h>
@@ -182,6 +183,26 @@ static unsigned int static_prio_timeslic
 }
 
 /*
+ * Divide a load by a sched group cpu_power : (load / sg->cpu_power)
+ * Since cpu_power is a 'constant', we can use a reciprocal divide.
+ */
+static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
+{
+	return reciprocal_divide(load, sg->reciprocal_cpu_power);
+}
+/*
+ * Each time a sched group cpu_power is changed,
+ * we must compute its reciprocal value
+ */
+static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
+{
+	sg->cpu_power += val;
+	BUG_ON(sg->cpu_power == 0);
+	sg->reciprocal_cpu_power = reciprocal_value(sg->cpu_power);
+}
+
+
+/*
  * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
  * to time slice values: [800ms ... 100ms ... 5ms]
  *
@@ -1241,7 +1262,8 @@ find_idlest_group(struct sched_domain *s
 		}
 
 		/* Adjust by relative CPU power of the group */
-		avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
+		avg_load = sg_div_cpu_power(group,
+				avg_load * SCHED_LOAD_SCALE);
 
 		if (local_group) {
 			this_load = avg_load;
@@ -2355,7 +2377,8 @@ find_busiest_group(struct sched_domain *
 		total_pwr += group->cpu_power;
 
 		/* Adjust by relative CPU power of the group */
-		avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
+		avg_load = sg_div_cpu_power(group,
+				avg_load * SCHED_LOAD_SCALE);
 
 		group_capacity = group->cpu_power / SCHED_LOAD_SCALE;
 
@@ -2510,8 +2533,8 @@ small_imbalance:
 		pwr_now /= SCHED_LOAD_SCALE;
 
 		/* Amount of load we'd subtract */
-		tmp = busiest_load_per_task * SCHED_LOAD_SCALE /
-			busiest->cpu_power;
+		tmp = sg_div_cpu_power(busiest,
+				busiest_load_per_task * SCHED_LOAD_SCALE);
 		if (max_load > tmp)
 			pwr_move += busiest->cpu_power *
 				min(busiest_load_per_task, max_load - tmp);
@@ -2519,10 +2542,11 @@ small_imbalance:
 		/* Amount of load we'd add */
 		if (max_load * busiest->cpu_power <
 				busiest_load_per_task * SCHED_LOAD_SCALE)
-			tmp = max_load * busiest->cpu_power / this->cpu_power;
+			tmp = sg_div_cpu_power(this,
+					max_load * busiest->cpu_power);
 		else
-			tmp = busiest_load_per_task * SCHED_LOAD_SCALE /
-				this->cpu_power;
+			tmp = sg_div_cpu_power(this,
+				busiest_load_per_task * SCHED_LOAD_SCALE);
 		pwr_move += this->cpu_power *
 			min(this_load_per_task, this_load + tmp);
 		pwr_move /= SCHED_LOAD_SCALE;
@@ -6352,7 +6376,7 @@ next_sg:
 			continue;
 		}
 
-		sg->cpu_power += sd->groups->cpu_power;
+		sg_inc_cpu_power(sg, sd->groups->cpu_power);
 	}
 	sg = sg->next;
 	if (sg != group_head)
@@ -6427,6 +6451,8 @@ static void init_sched_groups_power(int 
 
 	child = sd->child;
 
+	sd->groups->cpu_power = 0;
+
 	/*
 	 * For perf policy, if the groups in child domain share resources
 	 * (for example cores sharing some portions of the cache hierarchy
@@ -6437,18 +6463,16 @@ static void init_sched_groups_power(int 
 	if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
 		       (child->flags &
 			(SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
-		sd->groups->cpu_power = SCHED_LOAD_SCALE;
+		sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
 		return;
 	}
 
-	sd->groups->cpu_power = 0;
-
 	/*
 	 * add cpu_power of each child group to this groups cpu_power
 	 */
 	group = child->groups;
 	do {
-		sd->groups->cpu_power += group->cpu_power;
+		sg_inc_cpu_power(sd->groups, group->cpu_power);
 		group = group->next;
 	} while (group != child->groups);
 }