The PELT decay_load comments are a bit confusing, first of all
the 1/2^N should be (1/2)^N so that the reader doesn't get confused.
Secondly, the y^N splitting into a 2-part decay factor deserves
a better explanation. This patch improves the comments.
Cc: Paul Turner <[email protected]>
Cc: Dietmar Eggemann <[email protected]>
Cc: Juri Lelli <[email protected]>
Cc: Patrick Bellasi <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Signed-off-by: Joel Fernandes <[email protected]>
---
kernel/sched/fair.c | 14 +++++++++-----
1 file changed, 9 insertions(+), 5 deletions(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 6559d197e08a..1e1f2d77751e 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -2761,11 +2761,15 @@ static __always_inline u64 decay_load(u64 val, u64 n)
local_n = n;
/*
- * As y^PERIOD = 1/2, we can combine
- * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
- * With a look-up table which covers y^n (n<PERIOD)
- *
- * To achieve constant time decay_load.
+ * As y^PERIOD = 1/2, we can decay the load by 1/2
+ * for n/PERIOD number of PERIOD sized instances. Then
+ * we decay by the remaining windows in the final PERIOD,
+ * that is n%PERIOD. In other words, the decay factor
+ * y^N in terms of PERIOD becomes:
+ * y^n = (1/2)^(n/PERIOD) * y^(n%PERIOD).
+ * Since now we only need to compute powers of y where
+ * n < PERIOD, we use a look-up table for y^N (N<PERIOD).
+ * This helps achieve constant time decay_load.
*/
if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
val >>= local_n / LOAD_AVG_PERIOD;
--
2.12.0.246.ga2ecc84866-goog
On Fri, Mar 10, 2017 at 12:23:41PM -0800, Joel Fernandes wrote:
> The PELT decay_load comments are a bit confusing, first of all
> the 1/2^N should be (1/2)^N so that the reader doesn't get confused.
I'm thinking you're confused. They're identical.
(1/2)^N = (2^-1)^N = 2^-N = 1/2^N
> Secondly, the y^N splitting into a 2-part decay factor deserves
> a better explanation. This patch improves the comments.
I find its actually harder to read.
On Wed, Mar 22, 2017 at 7:16 AM, Peter Zijlstra <[email protected]> wrote:
> On Fri, Mar 10, 2017 at 12:23:41PM -0800, Joel Fernandes wrote:
>> The PELT decay_load comments are a bit confusing, first of all
>> the 1/2^N should be (1/2)^N so that the reader doesn't get confused.
>
> I'm thinking you're confused. They're identical.
>
> (1/2)^N = (2^-1)^N = 2^-N = 1/2^N
They are identical I know, but I meant by enclosing the 1/2 in
brackets, it is more clear that we multiply by 1/2 N times to the
first time reader - for the reason that we'd like to reduce the PELT
calculated load by 1/2 N times.
>> Secondly, the y^N splitting into a 2-part decay factor deserves
>> a better explanation. This patch improves the comments.
>
> I find its actually harder to read.
Oh, which part? Can you help improve it? Maybe I didn't word something
correctly?
Regards,
Joel
On Wed, Mar 22, 2017 at 09:35:43AM -0700, Joel Fernandes wrote:
> On Wed, Mar 22, 2017 at 7:16 AM, Peter Zijlstra <[email protected]> wrote:
> > On Fri, Mar 10, 2017 at 12:23:41PM -0800, Joel Fernandes wrote:
> >> The PELT decay_load comments are a bit confusing, first of all
> >> the 1/2^N should be (1/2)^N so that the reader doesn't get confused.
> >
> > I'm thinking you're confused. They're identical.
> >
> > (1/2)^N = (2^-1)^N = 2^-N = 1/2^N
>
> They are identical I know, but I meant by enclosing the 1/2 in
> brackets, it is more clear that we multiply by 1/2 N times to the
> first time reader - for the reason that we'd like to reduce the PELT
> calculated load by 1/2 N times.
Must be me then, because I've never been confused about that. Esp. so
since the first part: y^p = 1/2, explicitly mentions half. So its clear
from the factorization that half is meant.
> >> Secondly, the y^N splitting into a 2-part decay factor deserves
> >> a better explanation. This patch improves the comments.
> >
> > I find its actually harder to read.
>
> Oh, which part? Can you help improve it? Maybe I didn't word something
> correctly?
I think the fact that there's now words actually makes it worse.
The equation very concisely shows what we do. I don't see why we need
extra words there to obscure things.
Hi Peter,
On Wed, Mar 22, 2017 at 10:02 AM, Peter Zijlstra <[email protected]> wrote:
> On Wed, Mar 22, 2017 at 09:35:43AM -0700, Joel Fernandes wrote:
>> On Wed, Mar 22, 2017 at 7:16 AM, Peter Zijlstra <[email protected]> wrote:
>> > On Fri, Mar 10, 2017 at 12:23:41PM -0800, Joel Fernandes wrote:
>> >> The PELT decay_load comments are a bit confusing, first of all
>> >> the 1/2^N should be (1/2)^N so that the reader doesn't get confused.
>> >
>> > I'm thinking you're confused. They're identical.
>> >
>> > (1/2)^N = (2^-1)^N = 2^-N = 1/2^N
>>
>> They are identical I know, but I meant by enclosing the 1/2 in
>> brackets, it is more clear that we multiply by 1/2 N times to the
>> first time reader - for the reason that we'd like to reduce the PELT
>> calculated load by 1/2 N times.
>
> Must be me then, because I've never been confused about that. Esp. so
> since the first part: y^p = 1/2, explicitly mentions half. So its clear
> from the factorization that half is meant.
Yes that's true.
>> >> Secondly, the y^N splitting into a 2-part decay factor deserves
>> >> a better explanation. This patch improves the comments.
>> >
>> > I find its actually harder to read.
>>
>> Oh, which part? Can you help improve it? Maybe I didn't word something
>> correctly?
>
> I think the fact that there's now words actually makes it worse.
>
> The equation very concisely shows what we do. I don't see why we need
> extra words there to obscure things.
Ok, I agree with you and will kill this patch then. Thanks for the review.
Regards,
Joel