sched: smpnice try to wakeup modification
PATCH 1/1 sched: smpnice try to wakeup modification
Peter,
If there is no load on this_cpu, (i.e. tl_per_task is 0), we will fail the "tl + target_load(cpu, idx) <= tl_per_task" check. I think the original intention was to put task on this_cpu if it has no load and when there's already one task on cpu. This helps spread tasks out for low load condition.
Thanks.
Tim
kernel/sched.c +0 -0
@@ -1393,7 +1393,7 @@ static int try_to_wake_up(task_t *p, uns
if (this_sd->flags & SD_WAKE_AFFINE) {
unsigned long tl = this_load;- unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu);
+ unsigned long sl_per_task = cpu_avg_load_per_task(cpu);
/*
* If sync wakeup then subtract the (maximum possible)
@@ -1404,7 +1404,7 @@ static int try_to_wake_up(task_t *p, uns
tl -= current->load_weight;
if ((tl <= load &&
- tl + target_load(cpu, idx) <= tl_per_task) ||
+ tl + target_load(cpu, idx) <= sl_per_task) ||
100*(tl + p->load_weight) <= imbalance*load) {
/*
* This domain has SD_WAKE_AFFINE andQuoted 3 lines · Tim Chen
Peter,
If there is no load on
this_cpu, (i.e.tl_per_taskis 0), we will fail the "tl +target_load(cpu, idx)<=tl_per_task" check.
This isn't the case. If this_cpu is idle tl_per_task will be set to SCHED_LOAD_SCALE (see implementation of cpu_avg_load_per_task()) and that expression should succeed unless the value returned by target_load(cpu, idx) is bigger than SCHED_LOAD_SCALE. This is exactly the same as would have happened with the original code.
(BTW cpu_avg_load_per_task()'s original implementation would have had the effect you describe but it was modified when it was realized that it would break the code in a lot of places (not just here). The thinking now is that if there isn't enough data available to calculate the average load per task for a run queue then the correct value to use is the theoretical average i.e. SCHED_LOAD_SCALE.)
Quoted 24 lines
I think the
original intention was
to put task onthis_cpuif it has no load and when there's already one task on cpu. This helps spread tasks out for low load condition.Thanks.
Tim
Signed-off-by Tim Chen <[email protected]>---
linux-2.6.16-mm2-a/kernel/sched.c2006-03-28 16:00:37.091779904
-0800
+++linux-2.6.16-mm2-b/kernel/sched.c2006-03-28 16:09:08.237074008
-0800(quoted diff) +1
-1@@ -1393,7 +1393,7 @@ static int try_to_wake_up(task_t *p, uns if (this_sd->flags & SD_WAKE_AFFINE) { unsigned long tl = this_load; - unsigned long tl_per_task = cpu_avg_load_per_task(this_cpu); + unsigned long sl_per_task = cpu_avg_load_per_task(cpu); /* * If sync wakeup then subtract the (maximum possible)
Review static int try_to_wake_up(task_t *p, uns · line 1410
Quoted 8 lines
(quoted diff) +1
-1@@ -1404,7 +1404,7 @@ static int try_to_wake_up(task_t *p, uns tl -= current->load_weight; if ((tl <= load && - tl + target_load(cpu, idx) <= tl_per_task) || + tl + target_load(cpu, idx) <= sl_per_task) || 100*(tl + p->load_weight) <= imbalance*load) { /* * This domain has SD_WAKE_AFFINE and
Nevertheless, in some cases, it was difficult to decide which run queue's average load per task should be used to replace SCHED_LOAD_SCALE and a careful review of the code by those with expertise in this area would be appreciated as I may have made mistakes.
Peter
Signature · 4 lines
Peter Williams [email protected] "Learning, n. The kind of ignorance distinguishing the studious." -- Ambrose Bierce
Quoted 9 lines · Peter Williams
Quoted 3 lines · Tim Chen
Peter,
If there is no load on
this_cpu, (i.e.tl_per_taskis 0), we will fail the "tl +target_load(cpu, idx)<=tl_per_task" check.This isn't the case. If
this_cpuis idletl_per_taskwill be set toSCHED_LOAD_SCALE(see implementation ofcpu_avg_load_per_task()) and that expression should succeed unless the value returned bytarget_load(cpu, idx)is bigger thanSCHED_LOAD_SCALE. This is exactly the same as would have happened with the original code.(BTW cpu_avg_load_per_task()'s original implementation would have had the effect you describe but it was modified when it was realized that it would break the code in a lot of places (not just here). The thinking now is that if there isn't enough data available to calculate the average load per task for a run queue then the correct value to use is the theoretical average i.e.
SCHED_LOAD_SCALE.)
Ah, I was looking at the original smpnice patch. Thanks for clarifying.
Tim