[PATCH] sched/fair: refine maximum periods for decay_load

[PATCH] sched/fair: refine maximum periods for decay_load

[Leo Yan]

In current code, decay_load() will consider to set load value to zero
after passing 32*64 ms. So this means max_load * (0.5^64) ~= 0.

Kernel can support maximum number of processes and threads to 2^29 and
set task with highest priority with nice=-20 (weight = 88761). So in
worst case, one CPU may have maximum load value is:

    max_load = 2^29 * 88761 < 2^46

In theory after pass 46 periods we can ensure load value will be decayed
to zero. So this patch is to change maximum periods from 64 to 48.

Signed-off-by: Leo Yan <leo.yan@linaro.org>
---
 kernel/sched/fair.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index e342159..55cb134 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -2622,7 +2622,7 @@ static __always_inline u64 decay_load(u64 val, u64 n)
 
 	if (!n)
 		return val;
-	else if (unlikely(n > LOAD_AVG_PERIOD * 63))
+	else if (unlikely(n > LOAD_AVG_PERIOD * 47))
 		return 0;
 
 	/* after bounds checking we can collapse to 32-bit */
-- 
1.9.1

Re: [PATCH] sched/fair: refine maximum periods for decay_load

[Vincent Guittot]

Hi Leo,

On 5 August 2016 at 08:34, Leo Yan <leo.yan@linaro.org> wrote:
>
> In current code, decay_load() will consider to set load value to zero
> after passing 32*64 ms. So this means max_load * (0.5^64) ~= 0.
>
> Kernel can support maximum number of processes and threads to 2^29 and
> set task with highest priority with nice=-20 (weight = 88761). So in
> worst case, one CPU may have maximum load value is:
>
>     max_load = 2^29 * 88761 < 2^46
>
> In theory after pass 46 periods we can ensure load value will be decayed
> to zero. So this patch is to change maximum periods from 64 to 48.
>
> Signed-off-by: Leo Yan <leo.yan@linaro.org>
> ---
>  kernel/sched/fair.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index e342159..55cb134 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -2622,7 +2622,7 @@ static __always_inline u64 decay_load(u64 val, u64 n)
>
>         if (!n)
>                 return val;
> -       else if (unlikely(n > LOAD_AVG_PERIOD * 63))
> +       else if (unlikely(n > LOAD_AVG_PERIOD * 47))

In the equation above, you use 46 then you mentioned : "change maximum
periods from 64 to 48." and finally you use 47.

>                 return 0;
>
>         /* after bounds checking we can collapse to 32-bit */
> --
> 1.9.1
>

Re: [PATCH] sched/fair: refine maximum periods for decay_load

[Leo Yan]

Hi Vincent,

Thanks for review.

On Fri, Aug 05, 2016 at 09:56:46AM +0200, Vincent Guittot wrote:
> Hi Leo,
> 
> On 5 August 2016 at 08:34, Leo Yan <leo.yan@linaro.org> wrote:
> >
> > In current code, decay_load() will consider to set load value to zero
> > after passing 32*64 ms. So this means max_load * (0.5^64) ~= 0.
> >
> > Kernel can support maximum number of processes and threads to 2^29 and
> > set task with highest priority with nice=-20 (weight = 88761). So in
> > worst case, one CPU may have maximum load value is:
> >
> >     max_load = 2^29 * 88761 < 2^46
> >
> > In theory after pass 46 periods we can ensure load value will be decayed
> > to zero. So this patch is to change maximum periods from 64 to 48.
> >
> > Signed-off-by: Leo Yan <leo.yan@linaro.org>
> > ---
> >  kernel/sched/fair.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index e342159..55cb134 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -2622,7 +2622,7 @@ static __always_inline u64 decay_load(u64 val, u64 n)
> >
> >         if (!n)
> >                 return val;
> > -       else if (unlikely(n > LOAD_AVG_PERIOD * 63))
> > +       else if (unlikely(n > LOAD_AVG_PERIOD * 47))
> 
> In the equation above, you use 46 then you mentioned : "change maximum
> periods from 64 to 48." and finally you use 47.

Sorry introduce confusion. I want to align to 16 so choose maximum
periods to 48, and due to condition is (n > LOAD_AVG_PERIOD * 47) so
it equeals to (n >= LOAD_AVG_PERIOD * 48).

Precisely I think the code should be:
    else if (unlikely(n >= LOAD_AVG_PERIOD * 46))

Do you think this is okay?

> >                 return 0;
> >
> >         /* after bounds checking we can collapse to 32-bit */
> > --
> > 1.9.1
> >

Re: [PATCH] sched/fair: refine maximum periods for decay_load

[Vincent Guittot]

On 5 August 2016 at 10:17, Leo Yan <leo.yan@linaro.org> wrote:
> Hi Vincent,
>
> Thanks for review.
>
> On Fri, Aug 05, 2016 at 09:56:46AM +0200, Vincent Guittot wrote:
>> Hi Leo,
>>
>> On 5 August 2016 at 08:34, Leo Yan <leo.yan@linaro.org> wrote:
>> >
>> > In current code, decay_load() will consider to set load value to zero
>> > after passing 32*64 ms. So this means max_load * (0.5^64) ~= 0.
>> >
>> > Kernel can support maximum number of processes and threads to 2^29 and
>> > set task with highest priority with nice=-20 (weight = 88761). So in
>> > worst case, one CPU may have maximum load value is:
>> >
>> >     max_load = 2^29 * 88761 < 2^46
>> >
>> > In theory after pass 46 periods we can ensure load value will be decayed
>> > to zero. So this patch is to change maximum periods from 64 to 48.
>> >
>> > Signed-off-by: Leo Yan <leo.yan@linaro.org>
>> > ---
>> >  kernel/sched/fair.c | 2 +-
>> >  1 file changed, 1 insertion(+), 1 deletion(-)
>> >
>> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> > index e342159..55cb134 100644
>> > --- a/kernel/sched/fair.c
>> > +++ b/kernel/sched/fair.c
>> > @@ -2622,7 +2622,7 @@ static __always_inline u64 decay_load(u64 val, u64 n)
>> >
>> >         if (!n)
>> >                 return val;
>> > -       else if (unlikely(n > LOAD_AVG_PERIOD * 63))
>> > +       else if (unlikely(n > LOAD_AVG_PERIOD * 47))
>>
>> In the equation above, you use 46 then you mentioned : "change maximum
>> periods from 64 to 48." and finally you use 47.
>
> Sorry introduce confusion. I want to align to 16 so choose maximum
> periods to 48, and due to condition is (n > LOAD_AVG_PERIOD * 47) so
> it equeals to (n >= LOAD_AVG_PERIOD * 48).
>
> Precisely I think the code should be:
>     else if (unlikely(n >= LOAD_AVG_PERIOD * 46))
>
> Do you think this is okay?

yes, this looks okay


>
>> >                 return 0;
>> >
>> >         /* after bounds checking we can collapse to 32-bit */
>> > --
>> > 1.9.1
>> >