[PATCH] sched: HT optimisation
authorNick Piggin <nickpiggin@yahoo.com.au>
Sat, 10 Sep 2005 07:26:19 +0000 (00:26 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Sat, 10 Sep 2005 17:06:23 +0000 (10:06 -0700)
If an idle sibling of an HT queue encounters a busy sibling, then make
higher level load balancing of the non-idle variety.

Performance of multiprocessor HT systems with low numbers of tasks
(generally < number of virtual CPUs) can be significantly worse than the
exact same workloads when running in non-HT mode.  The reason is largely
due to poor scheduling behaviour.

This patch improves the situation, making the performance gap far less
significant on one problematic test case (tbench).

Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
kernel/sched.c

index 8535e5c68f5b687279eb791dd7c47a91658a99be..46fdd0bb1ed6c5f9aca38f103a05716189f9e48e 100644 (file)
@@ -1906,7 +1906,7 @@ out:
  */
 static struct sched_group *
 find_busiest_group(struct sched_domain *sd, int this_cpu,
-                  unsigned long *imbalance, enum idle_type idle)
+                  unsigned long *imbalance, enum idle_type idle, int *sd_idle)
 {
        struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
        unsigned long max_load, avg_load, total_load, this_load, total_pwr;
@@ -1931,6 +1931,9 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
                avg_load = 0;
 
                for_each_cpu_mask(i, group->cpumask) {
+                       if (*sd_idle && !idle_cpu(i))
+                               *sd_idle = 0;
+
                        /* Bias balancing toward cpus of our domain */
                        if (local_group)
                                load = target_load(i, load_idx);
@@ -2074,10 +2077,14 @@ static int load_balance(int this_cpu, runqueue_t *this_rq,
        unsigned long imbalance;
        int nr_moved, all_pinned = 0;
        int active_balance = 0;
+       int sd_idle = 0;
+
+       if (idle != NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER)
+               sd_idle = 1;
 
        schedstat_inc(sd, lb_cnt[idle]);
 
-       group = find_busiest_group(sd, this_cpu, &imbalance, idle);
+       group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle);
        if (!group) {
                schedstat_inc(sd, lb_nobusyg[idle]);
                goto out_balanced;
@@ -2150,6 +2157,8 @@ static int load_balance(int this_cpu, runqueue_t *this_rq,
                        sd->balance_interval *= 2;
        }
 
+       if (!nr_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER)
+               return -1;
        return nr_moved;
 
 out_balanced:
@@ -2161,6 +2170,8 @@ out_balanced:
                        (sd->balance_interval < sd->max_interval))
                sd->balance_interval *= 2;
 
+       if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
+               return -1;
        return 0;
 }
 
@@ -2178,9 +2189,13 @@ static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
        runqueue_t *busiest = NULL;
        unsigned long imbalance;
        int nr_moved = 0;
+       int sd_idle = 0;
+
+       if (sd->flags & SD_SHARE_CPUPOWER)
+               sd_idle = 1;
 
        schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
-       group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);
+       group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE, &sd_idle);
        if (!group) {
                schedstat_inc(sd, lb_nobusyg[NEWLY_IDLE]);
                goto out_balanced;
@@ -2205,15 +2220,19 @@ static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
                spin_unlock(&busiest->lock);
        }
 
-       if (!nr_moved)
+       if (!nr_moved) {
                schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
-       else
+               if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
+                       return -1;
+       } else
                sd->nr_balance_failed = 0;
 
        return nr_moved;
 
 out_balanced:
        schedstat_inc(sd, lb_balanced[NEWLY_IDLE]);
+       if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER)
+               return -1;
        sd->nr_balance_failed = 0;
        return 0;
 }
@@ -2338,7 +2357,10 @@ static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
 
                if (j - sd->last_balance >= interval) {
                        if (load_balance(this_cpu, this_rq, sd, idle)) {
-                               /* We've pulled tasks over so no longer idle */
+                               /* We've pulled tasks over so either we're no
+                                * longer idle, or one of our SMT siblings is
+                                * not idle.
+                                */
                                idle = NOT_IDLE;
                        }
                        sd->last_balance += interval;