[RFC] sched.c : procfs tunables

[RFC] sched.c : procfs tunables

[Al Boldi]

Proper scheduling in a multi-tasking environment is critical to the success 
of a desktop OS.  Linux, being mainly a server OS, is currently tuned to 
scheduling defaults that may be appropriate only for the server scenario.

To enable Linux to play an effective role on the desktop, a more flexible 
approach is necessary.  An approach that would allow the end-User the 
freedom to adjust the OS to the specific environment at hand.

So instead of forcing a one-size fits all approach on the end-User, would not 
exporting sched.c tunables to the procfs present a flexible approach to the 
scheduling dilemma?

All comments that have a vested interest in enabling Linux on the desktop are 
most welcome, even if they describe other/better/smarter approaches.

Thanks!

--
Al

Re: [RFC] sched.c : procfs tunables

[Con Kolivas]

On Saturday 01 April 2006 00:23, Al Boldi wrote:
> Proper scheduling in a multi-tasking environment is critical to the success
> of a desktop OS.  Linux, being mainly a server OS, is currently tuned to
> scheduling defaults that may be appropriate only for the server scenario.
>
> To enable Linux to play an effective role on the desktop, a more flexible
> approach is necessary.  An approach that would allow the end-User the
> freedom to adjust the OS to the specific environment at hand.
>
> So instead of forcing a one-size fits all approach on the end-User, would
> not exporting sched.c tunables to the procfs present a flexible approach to
> the scheduling dilemma?
>
> All comments that have a vested interest in enabling Linux on the desktop
> are most welcome, even if they describe other/better/smarter approaches.

None of the current "tunables" have easily understandable heuristics. Even 
those that appear to be obvious, like timselice, are not. While exporting 
tunables is not a bad idea, exporting tunables that noone understands is not 
really helpful. Even with heavy documentation, changes are not immediately 
predictable, and parts of the scheduler "know" about the default tuning 
values and they'd be broken by modifying them. Other scheduler designs, or 
more infrastructure on the current one (like what Mike's working on) might 
make some more obvious tunables. I've already discussed what I think in that 
regard too on a similar email thread. Exporting them also incurs a not 
insignificant cost.

Cheers,
Con

Re: [RFC] sched.c : procfs tunables

[Al Boldi]

Con Kolivas wrote:
> On Saturday 01 April 2006 00:23, Al Boldi wrote:
> > Proper scheduling in a multi-tasking environment is critical to the
> > success of a desktop OS.  Linux, being mainly a server OS, is currently
> > tuned to scheduling defaults that may be appropriate only for the server
> > scenario.
> >
> > To enable Linux to play an effective role on the desktop, a more
> > flexible approach is necessary.  An approach that would allow the
> > end-User the freedom to adjust the OS to the specific environment at
> > hand.
> >
> > So instead of forcing a one-size fits all approach on the end-User,
> > would not exporting sched.c tunables to the procfs present a flexible
> > approach to the scheduling dilemma?
> >
> > All comments that have a vested interest in enabling Linux on the
> > desktop are most welcome, even if they describe other/better/smarter
> > approaches.
>
> None of the current "tunables" have easily understandable heuristics. Even
> those that appear to be obvious, like timselice, are not. While exporting
> tunables is not a bad idea, exporting tunables that noone understands is
> not really helpful.

Couldn't this be fixed with an autotuning module based on cpu/mem/ctxt 
performance?

Mike Galbraith wrote:
> Nope, not the existing tunables anyway.  The full effect of even a tiny
> scheduler knob tweak is hard to predict even if you've studied the code
> carefully.  These knobs are just not generic enough to be exposed IMHO.

Are you implying that the code is built around these tunables rather than 
using them?

Thanks!

--
Al

Re: [RFC] sched.c : procfs tunables

[Con Kolivas]

On Monday 03 April 2006 21:59, Al Boldi wrote:
> Con Kolivas wrote:
> > None of the current "tunables" have easily understandable heuristics.
> > Even those that appear to be obvious, like timselice, are not. While
> > exporting tunables is not a bad idea, exporting tunables that noone
> > understands is not really helpful.
>
> Couldn't this be fixed with an autotuning module based on cpu/mem/ctxt
> performance?

You're assuming there is some meaningful relationship between changes in 
cpu/mem/ctxt performance and these tunables, which isn't the case. 
Furthermore if this was the case, noone understands it, can predict it or 
know how to tune it. Just saying "autotune it" doesn't really tell us how 
exactly the change those tunables in relation to the other variables. Since 
Mike and I understand them reasonably well I think we'd both agree that there 
is no meaningful association.

Cheers,
Con

Re: [RFC] sched.c : procfs tunables

[Al Boldi]

Con Kolivas wrote:
> On Monday 03 April 2006 21:59, Al Boldi wrote:
> > Con Kolivas wrote:
> > > None of the current "tunables" have easily understandable heuristics.
> > > Even those that appear to be obvious, like timselice, are not. While
> > > exporting tunables is not a bad idea, exporting tunables that noone
> > > understands is not really helpful.
> >
> > Couldn't this be fixed with an autotuning module based on cpu/mem/ctxt
> > performance?
>
> You're assuming there is some meaningful relationship between changes in
> cpu/mem/ctxt performance and these tunables, which isn't the case.
> Furthermore if this was the case, noone understands it, can predict it or
> know how to tune it. Just saying "autotune it" doesn't really tell us how
> exactly the change those tunables in relation to the other variables.
> Since Mike and I understand them reasonably well I think we'd both agree
> that there is no meaningful association.

After playing w/ these tunables it occurred to me that they are really only 
deadline limits, w/ a direct relation to cpu/mem/ctxt perf.

i.e timeslice=1 on i386sx means something other than timeslice=1 on amd64

It follows that w/o autotuning, the static default values have to be selected 
to allow for a large underlying perf range w/ a preference for the high 
range.  This is also the reason why 2.6 feels really crummy on low perf 
ranges.

Autotuning the default values would allow to tighten this range specific to 
the hw used, thus allowing for a smoother desktop experience.

Thanks!

--
Al

Re: [RFC] sched.c : procfs tunables

[Bill Davidsen]

Al Boldi wrote:

>Con Kolivas wrote:
>  
>
>>On Monday 03 April 2006 21:59, Al Boldi wrote:
>>    
>>
>>>Con Kolivas wrote:
>>>      
>>>
>>>>None of the current "tunables" have easily understandable heuristics.
>>>>Even those that appear to be obvious, like timselice, are not. While
>>>>exporting tunables is not a bad idea, exporting tunables that noone
>>>>understands is not really helpful.
>>>>        
>>>>
>>>Couldn't this be fixed with an autotuning module based on cpu/mem/ctxt
>>>performance?
>>>      
>>>
>>You're assuming there is some meaningful relationship between changes in
>>cpu/mem/ctxt performance and these tunables, which isn't the case.
>>Furthermore if this was the case, noone understands it, can predict it or
>>know how to tune it. Just saying "autotune it" doesn't really tell us how
>>exactly the change those tunables in relation to the other variables.
>>Since Mike and I understand them reasonably well I think we'd both agree
>>that there is no meaningful association.
>>    
>>
>
>After playing w/ these tunables it occurred to me that they are really only 
>deadline limits, w/ a direct relation to cpu/mem/ctxt perf.
>
>i.e timeslice=1 on i386sx means something other than timeslice=1 on amd64
>
>It follows that w/o autotuning, the static default values have to be selected 
>to allow for a large underlying perf range w/ a preference for the high 
>range.  This is also the reason why 2.6 feels really crummy on low perf 
>ranges.
>  
>
Actually the lower HZ has something to do with that, and tuning 
swappiness can also help a lot.

>Autotuning the default values would allow to tighten this range specific to 
>the hw used, thus allowing for a smoother desktop experience.
>
-- 

bill davidsen <davidsen@tmr.com>
  CTO TMR Associates, Inc
  Doing interesting things with small computers since 1979

Re: [RFC] sched.c : procfs tunables

[Mike Galbraith]

On Mon, 2006-04-03 at 14:59 +0300, Al Boldi wrote:
> Mike Galbraith wrote:
> > Nope, not the existing tunables anyway.  The full effect of even a tiny
> > scheduler knob tweak is hard to predict even if you've studied the code
> > carefully.  These knobs are just not generic enough to be exposed IMHO.
> 
> Are you implying that the code is built around these tunables rather than 
> using them?

I'm not implying anything, I'm merely stating my opinion: these knobs
have subtle effects which render them unsuitable for export.

	-Mike

Re: [RFC] sched.c : procfs tunables

[Mike Galbraith]

On Fri, 2006-03-31 at 17:23 +0300, Al Boldi wrote:
> Proper scheduling in a multi-tasking environment is critical to the success 
> of a desktop OS.  Linux, being mainly a server OS, is currently tuned to 
> scheduling defaults that may be appropriate only for the server scenario.
> 
> To enable Linux to play an effective role on the desktop, a more flexible 
> approach is necessary.  An approach that would allow the end-User the 
> freedom to adjust the OS to the specific environment at hand.
> 
> So instead of forcing a one-size fits all approach on the end-User, would not 
> exporting sched.c tunables to the procfs present a flexible approach to the 
> scheduling dilemma?

Nope, not the existing tunables anyway.  The full effect of even a tiny
scheduler knob tweak is hard to predict even if you've studied the code
carefully.  These knobs are just not generic enough to be exposed IMHO.

	-Mike

Re: [RFC] sched.c : procfs tunables

[Bill Davidsen]

Al Boldi wrote:

>Proper scheduling in a multi-tasking environment is critical to the success 
>of a desktop OS.  Linux, being mainly a server OS, is currently tuned to 
>scheduling defaults that may be appropriate only for the server scenario.
>  
>
I'm not sure I would agree about distribution kernels, and kernel.org 
kernels certainly have the options to trade overhead for more response.

>To enable Linux to play an effective role on the desktop, a more flexible 
>approach is necessary.  An approach that would allow the end-User the 
>freedom to adjust the OS to the specific environment at hand.
>
>So instead of forcing a one-size fits all approach on the end-User, would not 
>exporting sched.c tunables to the procfs present a flexible approach to the 
>scheduling dilemma?
>  
>

Let me agree with Mike and Con, I understand just well enough to pretty 
much leave them alone. The swappiness is available, that's one of the 
things which wants tuning. But the old 2.2 kernels did run better on 
small machines, even a stripped 2.6 kernel is slower.

>All comments that have a vested interest in enabling Linux on the desktop are 
>most welcome, even if they describe other/better/smarter approaches.
>
>Thanks!
>
>--
>Al
>
>-
>To unsubscribe from this list: send the line "unsubscribe linux-smp" in
>the body of a message to majordomo@vger.kernel.org
>More majordomo info at  http://vger.kernel.org/majordomo-info.html
>
>  
>


-- 
bill davidsen <davidsen@tmr.com>
  CTO TMR Associates, Inc
  Doing interesting things with small computers since 1979