[PATCH 3/3] sched: scale cpu load for judgment of group imbalance

From: Lei Wen
Date: Mon Jun 17 2013 - 09:02:56 EST


We cannot compare two load directly from two cpus, since the cpu power
over two cpu may vary largely.

Suppose we meet such two kind of cpus.
CPU A:
No real time work, and there are 3 task, with rq->load.weight
being 512.
CPU B:
Has real time work, and it take 3/4 of the cpu power, which
makes CFS only take 1/4, that is 1024/4=256 cpu power. And over its CFS
runqueue, there is only one task with weight as 128.

Since both cpu's CFS task take for half of the CFS's cpu power, it
should be considered as balanced in such case.

But original judgment like:
if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
(max_nr_running - min_nr_running) > 1)
It makes (512-128)>=((512+128)/4), and lead to imbalance conclusion...

Make the load as scaled, to avoid such case.

Signed-off-by: Lei Wen <leiwen@xxxxxxxxxxx>
---
kernel/sched/fair.c | 19 ++++++++++++-------
1 file changed, 12 insertions(+), 7 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 77a149c..18625dd 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -4434,7 +4434,7 @@ static inline void update_sg_lb_stats(struct lb_env *env,
int local_group, int *balance, struct sg_lb_stats *sgs)
{
unsigned long nr_running, max_nr_running, min_nr_running;
- unsigned long load, max_cpu_load, min_cpu_load;
+ unsigned long scaled_load, load, max_cpu_load, min_cpu_load;
unsigned int balance_cpu = -1, first_idle_cpu = 0;
unsigned long avg_load_per_task = 0;
int i;
@@ -4464,10 +4464,12 @@ static inline void update_sg_lb_stats(struct lb_env *env,
load = target_load(i, load_idx);
} else {
load = source_load(i, load_idx);
- if (load > max_cpu_load)
- max_cpu_load = load;
- if (min_cpu_load > load)
- min_cpu_load = load;
+ scaled_load = load * SCHED_POWER_SCALE
+ / cpu_rq(i)->cpu_power;
+ if (scaled_load > max_cpu_load)
+ max_cpu_load = scaled_load;
+ if (min_cpu_load > scaled_load)
+ min_cpu_load = scaled_load;

if (nr_running > max_nr_running)
max_nr_running = nr_running;
@@ -4511,8 +4513,11 @@ static inline void update_sg_lb_stats(struct lb_env *env,
* normalized nr_running number somewhere that negates
* the hierarchy?
*/
- if (sgs->sum_nr_running)
- avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
+ if (sgs->sum_nr_running) {
+ avg_load_per_task = sgs->sum_weighted_load * SCHED_POWER_SCALE
+ / group->sgp->power;
+ avg_load_per_task /= sgs->sum_nr_running;
+ }

if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
(max_nr_running - min_nr_running) > 1)
--
1.7.10.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/