Problems caused by scheduler tweaks in 2.6.0-test6?

Problems caused by scheduler tweaks in 2.6.0-test6?

[Tom Sightler]

Hi All,

Over the last few months I have tested many different scheduler tweaks
mostly by testing the -mm kernels and also by applying Nick's patches
against vanilla kernels.  Up until recently I have been very happy with
2.6.0-test5 with Nick's scheduler patches.

Then I decided to try 2.6.0-test6 which seems to include a lot of Con's
work and, while overall this seems nice, I'm having two relatively
serious side effects that seem to be related to this inclusion.

1.  VMware performance varies wildly.  I can't put my finger on this
exact issue, but I have found as way to repeatably trigger bad
performance.  When running VMware in fullscreen mode, enable window
animation and repeatedly minimize/maximize a window.  Under 2.4.x and
2.6.0-test5 w/Nick's patches this process runs reasonably smooth,
although noticably slower than native speed.  With stock 2.6.0-test6
after only a few seconds the minimize/maximize animiation slows to a
complete crawl, take 20+ seconds to complete the minimize opertaion. 
I've tried tuning VMware with priorities but no luck.

2.  I also use Wine to run various Windows programs on occasion,
particularly Outlook 2000 (mainly when attempting to help other running
this application on Windows).  The program runs fine, but always hangs
on exit.  I didn't originally think this was related to the scheduler,
but interestingly, after applying Nick's patches to 2.6.0-test6, which
back out Con's changes, this problem goes away.

Is there any help out there for these type of issues?  I know that many
people seem to think these changes make life better, and I'll admidt
that playing MP3's and DVD's is better with these changes, but I'd
rather have my system preform well at other tasks.  I would think having
a way to turn off all the fancy interactivity detection would be ideal
but there always seems to be opposition to adding tuning knobs.

Later,
Tom

Re: Problems caused by scheduler tweaks in 2.6.0-test6?

[Nick Piggin]

Did you see this Con?

By the way Tom, I have my scheduler patch available for test6 here:
http://www.kerneltrap.org/~npiggin/v15a/sched-rollup-v15a-260t6.gz

Tom Sightler wrote:

>Hi All,
>
>Over the last few months I have tested many different scheduler tweaks
>mostly by testing the -mm kernels and also by applying Nick's patches
>against vanilla kernels.  Up until recently I have been very happy with
>2.6.0-test5 with Nick's scheduler patches.
>
>Then I decided to try 2.6.0-test6 which seems to include a lot of Con's
>work and, while overall this seems nice, I'm having two relatively
>serious side effects that seem to be related to this inclusion.
>
>1.  VMware performance varies wildly.  I can't put my finger on this
>exact issue, but I have found as way to repeatably trigger bad
>performance.  When running VMware in fullscreen mode, enable window
>animation and repeatedly minimize/maximize a window.  Under 2.4.x and
>2.6.0-test5 w/Nick's patches this process runs reasonably smooth,
>although noticably slower than native speed.  With stock 2.6.0-test6
>after only a few seconds the minimize/maximize animiation slows to a
>complete crawl, take 20+ seconds to complete the minimize opertaion. 
>I've tried tuning VMware with priorities but no luck.
>
>2.  I also use Wine to run various Windows programs on occasion,
>particularly Outlook 2000 (mainly when attempting to help other running
>this application on Windows).  The program runs fine, but always hangs
>on exit.  I didn't originally think this was related to the scheduler,
>but interestingly, after applying Nick's patches to 2.6.0-test6, which
>back out Con's changes, this problem goes away.
>
>Is there any help out there for these type of issues?  I know that many
>people seem to think these changes make life better, and I'll admidt
>that playing MP3's and DVD's is better with these changes, but I'd
>rather have my system preform well at other tasks.  I would think having
>a way to turn off all the fancy interactivity detection would be ideal
>but there always seems to be opposition to adding tuning knobs.
>
>Later,
>Tom
>
>  
>

Re: Problems caused by scheduler tweaks in 2.6.0-test6?

[Con Kolivas]

On Sat, 4 Oct 2003 19:11, Nick Piggin wrote:
> Did you see this Con?
>
> By the way Tom, I have my scheduler patch available for test6 here:
> http://www.kerneltrap.org/~npiggin/v15a/sched-rollup-v15a-260t6.gz
>
> Tom Sightler wrote:
> >Hi All,
> >
> >Over the last few months I have tested many different scheduler tweaks
> >mostly by testing the -mm kernels and also by applying Nick's patches
> >against vanilla kernels.  Up until recently I have been very happy with
> >2.6.0-test5 with Nick's scheduler patches.
> >
> >Then I decided to try 2.6.0-test6 which seems to include a lot of Con's
> >work and, while overall this seems nice, I'm having two relatively
> >serious side effects that seem to be related to this inclusion.
> >
> >1.  VMware performance varies wildly.  I can't put my finger on this
> >exact issue, but I have found as way to repeatably trigger bad
> >performance.  When running VMware in fullscreen mode, enable window
> >animation and repeatedly minimize/maximize a window.  Under 2.4.x and
> >2.6.0-test5 w/Nick's patches this process runs reasonably smooth,
> >although noticably slower than native speed.  With stock 2.6.0-test6
> >after only a few seconds the minimize/maximize animiation slows to a
> >complete crawl, take 20+ seconds to complete the minimize opertaion.
> >I've tried tuning VMware with priorities but no luck.
> >
> >2.  I also use Wine to run various Windows programs on occasion,
> >particularly Outlook 2000 (mainly when attempting to help other running
> >this application on Windows).  The program runs fine, but always hangs
> >on exit.  I didn't originally think this was related to the scheduler,
> >but interestingly, after applying Nick's patches to 2.6.0-test6, which
> >back out Con's changes, this problem goes away.
> >
> >Is there any help out there for these type of issues?  I know that many
> >people seem to think these changes make life better, and I'll admidt
> >that playing MP3's and DVD's is better with these changes, but I'd
> >rather have my system preform well at other tasks.  I would think having
> >a way to turn off all the fancy interactivity detection would be ideal
> >but there always seems to be opposition to adding tuning knobs.
> >

Please send a rundown of what top shows during these occurrences, and please 
define "hangs". I can't see how the scheduler tweaks can bring the machine 
down.

Con

Re: Problems caused by scheduler tweaks in 2.6.0-test6?

[Nick Piggin]

Con Kolivas wrote:

>On Sat, 4 Oct 2003 19:11, Nick Piggin wrote:
>
>>Did you see this Con?
>>
>>By the way Tom, I have my scheduler patch available for test6 here:
>>http://www.kerneltrap.org/~npiggin/v15a/sched-rollup-v15a-260t6.gz
>>
>>Tom Sightler wrote:
>>
>>>Hi All,
>>>
>>>Over the last few months I have tested many different scheduler tweaks
>>>mostly by testing the -mm kernels and also by applying Nick's patches
>>>against vanilla kernels.  Up until recently I have been very happy with
>>>2.6.0-test5 with Nick's scheduler patches.
>>>
>>>Then I decided to try 2.6.0-test6 which seems to include a lot of Con's
>>>work and, while overall this seems nice, I'm having two relatively
>>>serious side effects that seem to be related to this inclusion.
>>>
>>>1.  VMware performance varies wildly.  I can't put my finger on this
>>>exact issue, but I have found as way to repeatably trigger bad
>>>performance.  When running VMware in fullscreen mode, enable window
>>>animation and repeatedly minimize/maximize a window.  Under 2.4.x and
>>>2.6.0-test5 w/Nick's patches this process runs reasonably smooth,
>>>although noticably slower than native speed.  With stock 2.6.0-test6
>>>after only a few seconds the minimize/maximize animiation slows to a
>>>complete crawl, take 20+ seconds to complete the minimize opertaion.
>>>I've tried tuning VMware with priorities but no luck.
>>>
>>>2.  I also use Wine to run various Windows programs on occasion,
>>>particularly Outlook 2000 (mainly when attempting to help other running
>>>this application on Windows).  The program runs fine, but always hangs
>>>on exit.  I didn't originally think this was related to the scheduler,
>>>but interestingly, after applying Nick's patches to 2.6.0-test6, which
>>>back out Con's changes, this problem goes away.
>>>
>>>Is there any help out there for these type of issues?  I know that many
>>>people seem to think these changes make life better, and I'll admidt
>>>that playing MP3's and DVD's is better with these changes, but I'd
>>>rather have my system preform well at other tasks.  I would think having
>>>a way to turn off all the fancy interactivity detection would be ideal
>>>but there always seems to be opposition to adding tuning knobs.
>>>
>>>
>
>Please send a rundown of what top shows during these occurrences, and please 
>define "hangs". I can't see how the scheduler tweaks can bring the machine 
>down.
>

Could be unrelated. Might be a livelock or priority inversion spinning 
thingy

Re: Problems caused by scheduler tweaks in 2.6.0-test6?

[Tom Sightler]

> Please send a rundown of what top shows during these occurrences, and please 
> define "hangs". I can't see how the scheduler tweaks can bring the machine 
> down.
> 
> Con

I'm not sure exactly how to get the output from top when running VMware
fullscreen, does it have a way to dump the output to text?  Would
another tool provide you with valuable output?

As far as the Wine/Outlook 2000 issue, I don't know how the scheduler
changes cause this, and for the longest time I never really thought that
was it, I always assumed that it was something else different between
the main and the -mm tree.  However, I'm basically 100% sure that it's
the cause now because I tested with stock -test6 and then applied Nick's
patches for -test6 (which back out your changes) and the problem goes
away.

Hang means that the program simply will never exit, not that the whole
system hangs or anything.  When you exit Outlook it synchronizes the
mailbox before it exits and the closes down the connection.  The program
hangs with a "Please wait while Outlook exits" message a this never goes
away.  It's 100% repeatable with your scheduler tweaks, and never
happens with Nick's patches.  What can I provide for that issue?  Top
output shows the process sleeping, not using any CPU.  I can kill the
process and it will exit.

I guess my concern is that I seem to run into a lot of issues that I
consider bad behavior.  I've outlined the issues with Acrobat to you
before and you stated that you had profiled it.  Well this is something
I use almost every day and with these scheduler tweaks it becomes almost
unusable.  With Nicks patches, or even stock -test5 it works great.

Basically for me these tweaks significantly degrade three major programs
that I use to do work related activities (VMware, Acrobat, and to some
extent Crossover Office).  I'll be glad to help where I can, tell me
what information would help you.

Later,
Tom