According to the comment for post_init_entity_util_avg(), it seems that
we are assuming se->load.weight has the same scale/unit as that of
cfs_rq->avg.load_avg.
As far as I understand, se->load.weight is the scaled-up load, instead
of the true weight (as mapped directly from the nice value) of a task.
When CONFIG_32BIT is set, we have load == weight; when CONFIG_64BIT is
set, we have load == weight * 1024. However, cfs_rq->avg.load_avg is
the sum of true weights of tasks, as se->avg.load_avg corresponds to
the true weight of a task.
Based on how sa->util_avg is calculated in the code, we could be
inflating sa->util_avg by 1024 times? Could this be the reason why
"However, in many cases, the above util_avg does not give a desired
value. ... "?
I'm not entirely sure about it. Posting this for clarification and
comments.
Signed-off-by: Dawei Li <[email protected]>
---
kernel/sched/fair.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 6a16129f9a5c..0d13e52e1a92 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -1031,7 +1031,8 @@ void init_entity_runnable_average(struct sched_entity *se)
* With new tasks being created, their initial util_avgs are extrapolated
* based on the cfs_rq's current util_avg:
*
- * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
+ * util_avg = cfs_rq->avg.util_avg / (cfs_rq->avg.load_avg + 1)
+ * * se_weight(se)
*
* However, in many cases, the above util_avg does not give a desired
* value. Moreover, the sum of the util_avgs may be divergent, such
@@ -1078,7 +1079,7 @@ void post_init_entity_util_avg(struct task_struct *p)
if (cap > 0) {
if (cfs_rq->avg.util_avg != 0) {
- sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
+ sa->util_avg = cfs_rq->avg.util_avg * se_weight(se);
sa->util_avg /= (cfs_rq->avg.load_avg + 1);
if (sa->util_avg > cap)
--
2.40.1
On Fri, 8 Mar 2024 at 21:23, Dawei Li <[email protected]> wrote:
>
> According to the comment for post_init_entity_util_avg(), it seems that
> we are assuming se->load.weight has the same scale/unit as that of
> cfs_rq->avg.load_avg.
>
> As far as I understand, se->load.weight is the scaled-up load, instead
> of the true weight (as mapped directly from the nice value) of a task.
> When CONFIG_32BIT is set, we have load == weight; when CONFIG_64BIT is
> set, we have load == weight * 1024. However, cfs_rq->avg.load_avg is
> the sum of true weights of tasks, as se->avg.load_avg corresponds to
> the true weight of a task.
>
> Based on how sa->util_avg is calculated in the code, we could be
> inflating sa->util_avg by 1024 times? Could this be the reason why
> "However, in many cases, the above util_avg does not give a desired
> value. ... "?
No, this is about the fact that this estimation of util_avg can give
insane value because even the correct se_weight(se) can go up to 88761
whereas util_avg should never go above 1024
>
> I'm not entirely sure about it. Posting this for clarification and
> comments.
Good catch.
The commit message needs to be updated to remove your question above
but otherwise looks good
>
> Signed-off-by: Dawei Li <[email protected]>
> ---
> kernel/sched/fair.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 6a16129f9a5c..0d13e52e1a92 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -1031,7 +1031,8 @@ void init_entity_runnable_average(struct sched_entity *se)
> * With new tasks being created, their initial util_avgs are extrapolated
> * based on the cfs_rq's current util_avg:
> *
> - * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
> + * util_avg = cfs_rq->avg.util_avg / (cfs_rq->avg.load_avg + 1)
> + * * se_weight(se)
> *
> * However, in many cases, the above util_avg does not give a desired
> * value. Moreover, the sum of the util_avgs may be divergent, such
> @@ -1078,7 +1079,7 @@ void post_init_entity_util_avg(struct task_struct *p)
>
> if (cap > 0) {
> if (cfs_rq->avg.util_avg != 0) {
> - sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
> + sa->util_avg = cfs_rq->avg.util_avg * se_weight(se);
> sa->util_avg /= (cfs_rq->avg.load_avg + 1);
>
> if (sa->util_avg > cap)
> --
> 2.40.1
>