performance -cfs19 vs. -ck1 for 2.6.22

performance -cfs19 vs. -ck1 for 2.6.22

[Ni@m]

Hi,
I haven't made special tests to compare 2.6.22-cfs19 and 2.6.22-ck1,
but I see performance superiority of 2.6.22-ck1 using my own eyes.
Especially in mozilla-firefox ...
Using 2.6.22 or 2.6.22-cfs19 my firefox hangs for some time(~4-5
seconds) when I'm trying to open any url, but with 2.6.22-ck1 firefox
doesn't hang at all or at least for a second when I have lots of open
tabs.

What is the reason? Had Con made better scheduler than cfs and O(1) in
general kernel?

Kernel configuration in 2.6.22, 2.6.22-ck1, 2.6.22-cfs19 is similar.

Re: performance -cfs19 vs. -ck1 for 2.6.22

[Ingo Molnar]

* Ni@m <niam.niam@gmail.com> wrote:

> Using 2.6.22 or 2.6.22-cfs19 my firefox hangs for some time(~4-5 
> seconds) when I'm trying to open any url, but with 2.6.22-ck1 firefox 
> doesn't hang at all or at least for a second when I have lots of open 
> tabs.

hm, is there no other workload on the system? Could you run the 
following script:

   http://people.redhat.com/mingo/cfs-scheduler/tools/cfs-debug-info.sh

while you are busy opening new tabs on a CFSv19 kernel (and send me the 
file it outputs)?

One way to start up the script precisely would be to start it in a 
terminal like this:

  sleep 10; echo "soon ..."; sleep 1; ./cfs-debug-info.sh

that way you have 10 seconds to switch from the terminal to firefox, and 
you can time the delay of firefox precisely on the start-time of the 
debug script.

	Ingo

Re: performance -cfs19 vs. -ck1 for 2.6.22

[Ni@m]

Hi, Ingo!

>>Could you run the
>>following script:
>>  http://people.redhat.com/mingo/cfs-scheduler/tools/cfs-debug-info.sh
I'll do that! But a little bit later - I have tested that on my home computer.

>>hm, is there no other workload on the system?
Workload: firefox, transmission(torrent client), gnome terminal, music player.

Now I think it could be not a scheduler issue but network subsystem
... but I've got such situation on previous kernels and w/o torrent
client running.
The problem is not in switching tabs in firefox but in _opening_new_url_.
I'm not good in Con's patches but are they touching network subsystem?

Re: performance -cfs19 vs. -ck1 for 2.6.22

[Ingo Molnar]

* Ni@m <niam.niam@gmail.com> wrote:

> Hi, Ingo!
> 
> >>Could you run the
> >>following script:
> >> http://people.redhat.com/mingo/cfs-scheduler/tools/cfs-debug-info.sh
> I'll do that! But a little bit later - I have tested that on my home 
> computer.
> 
> >>hm, is there no other workload on the system?
> Workload: firefox, transmission(torrent client), gnome terminal, music 
> player.
> 
> Now I think it could be not a scheduler issue but network subsystem 
> ... but I've got such situation on previous kernels and w/o torrent 
> client running.
>
> The problem is not in switching tabs in firefox but in 
> _opening_new_url_. I'm not good in Con's patches but are they touching 
> network subsystem?

Con's patches are not really touching the network subsystem, but we want 
to analyze this and figure out where the difference in behavior comes 
from, be that scheduling, networking or something else.

to help us figure out the nature of the delay, could you do another 
thing as well:

  strace -ttt -TTT -f -o firefox.trace.txt -p `pidof firefox-bin`

this will produce firefox.trace.txt - compress that via bzip -9. (and 
send that file privately - it might include private information embedded 
in URLs, etc.) Please do such a trace both under CFS and under -ck, 
under similar circumstances.

To isolate it down to the CPU scheduler changes you can create a pure 
2.6.22+SD kernel by applying this patch to a vanilla v2.6.22 tree:

  http://kernel.org/pub/linux/kernel/people/ck/patches/2.6/2.6.22/2.6.22-ck1/patches/sched-staircase-deadline-cpu-scheduler.patch

	Ingo

Fwd: performance -cfs19 vs. -ck1 for 2.6.22

[Ni@m]

> to help us figure out the nature of the delay, could you do another
> thing as well:
>
>   strace -ttt -TTT -f -o firefox.trace.txt -p `pidof firefox-bin`
>

I'll do that!!