Poor performance now? Please, put weighted velocities ctxts/s to the sched.decision!

Poor performance now? Please, put weighted velocities ctxts/s to the sched.decision!

[J.C. Pizarro]

[1] "Performance changes between 2.6.13 and 2.6.23"
http://lkml.org/lkml/2008/3/25/102
On Tue, Mar 25, 2008 at 1:34 PM, Sanders, Rob M. <sanders-rob@zai.com> wrote:
> Of more particular concern, the main application that I build has seen a 4-5X
> slowdown in performance.  Under YDL4.0.91 I could process roughly 1e8 data
> points in ~2 seconds, and under YDL6.0 I now process 1e8 data points in
> ~8 seconds.  Total CPU loading (from top) is about 5%, under both systems.
...
> YDL4.0.91 system is context switching about 7000 per interval, whereas the
> YDl6 system is context switching about 1200 times per interval.

It's the velocity, ctxts/interval, a parameter not recognized by the scheduler.

I said them to put the proposal of putting the counters and velocities fields
to the task's struct for the scheduler can take a good decision, the past month.

More weight in higher ctxts/s for the scheduler=> more interactivity.

[2] "Re: Serious performance regression in Wine applications and Linux 2.6.24.*"
http://lkml.org/lkml/2008/3/25/181
On Tue, 25 Mar 2008 08:49:13 -0700, "Ray Lee" wrote:
> Andi's idea of looking for excessive context switches is good -- I
> didn't see a response to that one. Other than that, if you're only
> noticing the issue in 3d games, then it could be several things (not
> just the scheduler).

The kernel needed the velocity parameter to gain it: ctxts per interval
(normally ctxts/s or ctxts/minute).

The proposal of the past month was here:

"Please, put 64-bit counter per task and incr.by.one each ctxt switch."
http://lkml.org/lkml/2008/2/23/398
http://lkml.org/lkml/2008/2/23/401
http://lkml.org/lkml/2008/2/23/406
http://lkml.org/lkml/2008/2/24/1
http://lkml.org/lkml/2008/2/24/103
http://lkml.org/lkml/2008/2/24/157
http://lkml.org/lkml/2008/2/26/159

My proposal is free for yours, quasi-bastards, I won't charge you!

   J.C.Pizarro

Re: Poor performance now? Please, put weighted velocities ctxts/s to the sched.decision!

[Ray Lee]

On Tue, Mar 25, 2008 at 1:29 PM, J.C. Pizarro <jcpiza@gmail.com> wrote:
>  [2] "Re: Serious performance regression in Wine applications and Linux 2.6.24.*"
>  http://lkml.org/lkml/2008/3/25/181
>  On Tue, 25 Mar 2008 08:49:13 -0700, "Ray Lee" wrote:
>  > Andi's idea of looking for excessive context switches is good -- I
>  > didn't see a response to that one. Other than that, if you're only
>  > noticing the issue in 3d games, then it could be several things (not
>  > just the scheduler).
>
>  The kernel needed the velocity parameter to gain it: ctxts per interval
>  (normally ctxts/s or ctxts/minute).

We don't know that's the problem yet. The reporter is doing more testing.

>  "Please, put 64-bit counter per task and incr.by.one each ctxt switch."
>  http://lkml.org/lkml/2008/2/23/398
>  http://lkml.org/lkml/2008/2/23/401
>  http://lkml.org/lkml/2008/2/23/406
>  http://lkml.org/lkml/2008/2/24/1
>  http://lkml.org/lkml/2008/2/24/103
>  http://lkml.org/lkml/2008/2/24/157
>  http://lkml.org/lkml/2008/2/26/159

There are already counters recording the switches:
    http://lkml.org/lkml/2008/2/25/314

>  My proposal is free for yours, quasi-bastards, I won't charge you!

My parents were married.

Re: Poor performance now? Please, put weighted velocities ctxts/s to the sched.decision!

[J.C. Pizarro]

On 2008/3/25, Ray Lee <ray-lk@madrabbit.org> wrote:
> On Tue, Mar 25, 2008 at 1:29 PM, J.C. Pizarro <jcpiza@gmail.com> wrote:
>  >  [2] "Re: Serious performance regression in Wine applications and Linux 2.6.24.*"
>  >  http://lkml.org/lkml/2008/3/25/181
>  >  On Tue, 25 Mar 2008 08:49:13 -0700, "Ray Lee" wrote:
>  >  > Andi's idea of looking for excessive context switches is good -- I
>  >  > didn't see a response to that one. Other than that, if you're only
>  >  > noticing the issue in 3d games, then it could be several things (not
>  >  > just the scheduler).
>  >
>  >  The kernel needed the velocity parameter to gain it: ctxts per interval
>  >  (normally ctxts/s or ctxts/minute).
>
>
> We don't know that's the problem yet. The reporter is doing more testing.
>
>
>  >  "Please, put 64-bit counter per task and incr.by.one each ctxt switch."
>  >  http://lkml.org/lkml/2008/2/23/398
>  >  http://lkml.org/lkml/2008/2/23/401
>  >  http://lkml.org/lkml/2008/2/23/406
>  >  http://lkml.org/lkml/2008/2/24/1
>  >  http://lkml.org/lkml/2008/2/24/103
>  >  http://lkml.org/lkml/2008/2/24/157
>  >  http://lkml.org/lkml/2008/2/26/159
>
>
> There are already counters recording the switches:
>     http://lkml.org/lkml/2008/2/25/314

I forgot this link, thanks!

It seems that they don't make ++ of counters in each switch ocurr.
I don't know how they record the switches.

The current scheduler don't take the count or the velocity of ctxts
as an extra parameter more in the decision of scheduling.

>  >  My proposal is free for yours, quasi-bastards, I won't charge you!
>
>
> My parents were married.

Luck boy ;) !