hackbench regression with 2.6.26-rc2 on tulsa machine

hackbench regression with 2.6.26-rc2 on tulsa machine

[Zhang, Yanmin]

Comparing with 2.6.26-rc1, hackbench has about 30% regression with 2.6.26-rc2 on my tulsa machine
which is a netburst architecure hyper-threading x86_64.

Command line to test: #hackbench 100 process 2000

With 2.6.26-rc1, it takes 30 seconds. But with 2.6.26-rc2/rc3, it takes 40 seconds.

Bisect located below patch:
46151122e0a2e80e5a6b2889f595e371fe2b600d is first bad commit
commit 46151122e0a2e80e5a6b2889f595e371fe2b600d
Author: Mike Galbraith <efault@gmx.de>
Date:   Thu May 8 17:00:42 2008 +0200

    sched: fix weight calculations
    
    The conversion between virtual and real time is as follows:
    
      dvt = rw/w * dt <=> dt = w/rw * dvt
    
    Since we want the fair sleeper granularity to be in real time, we actually
    need to do:
    
      dvt = - rw/w * l



The bisect steps look stable.

On my core architecure machines(stoakley and tigerton), I do see improvement instead of regression,
like result from 31 seconds down to 28 seconds.

I'm not sure if hyper-threading need more cares in this situation.

-yanmin

Re: hackbench regression with 2.6.26-rc2 on tulsa machine

[Mike Galbraith]

On Tue, 2008-05-20 at 16:09 +0800, Zhang, Yanmin wrote:
> Comparing with 2.6.26-rc1, hackbench has about 30% regression with 2.6.26-rc2 on my tulsa machine
> which is a netburst architecure hyper-threading x86_64.
> 
> Command line to test: #hackbench 100 process 2000
> 
> With 2.6.26-rc1, it takes 30 seconds. But with 2.6.26-rc2/rc3, it takes 40 seconds.
> 
> Bisect located below patch:
> 46151122e0a2e80e5a6b2889f595e371fe2b600d is first bad commit
> commit 46151122e0a2e80e5a6b2889f595e371fe2b600d
> Author: Mike Galbraith <efault@gmx.de>
> Date:   Thu May 8 17:00:42 2008 +0200
> 
>     sched: fix weight calculations
>     
>     The conversion between virtual and real time is as follows:
>     
>       dvt = rw/w * dt <=> dt = w/rw * dvt
>     
>     Since we want the fair sleeper granularity to be in real time, we actually
>     need to do:
>     
>       dvt = - rw/w * l
> 
> 
> 
> The bisect steps look stable.
> 
> On my core architecure machines(stoakley and tigerton), I do see improvement instead of regression,
> like result from 31 seconds down to 28 seconds.
> 
> I'm not sure if hyper-threading need more cares in this situation.

Oh joy.  I'll update my poor old P4 and see what I can duplicate this.

Do you still have group scheduling enabled?  If so, can you turn it off
and try again?  (when in doubt, grasp at any straw within reach;)

	-Mike

Re: hackbench regression with 2.6.26-rc2 on tulsa machine

[Zhang, Yanmin]

On Tue, 2008-05-20 at 11:22 +0200, Mike Galbraith wrote:
> On Tue, 2008-05-20 at 16:09 +0800, Zhang, Yanmin wrote:
> > Comparing with 2.6.26-rc1, hackbench has about 30% regression with 2.6.26-rc2 on my tulsa machine
> > which is a netburst architecure hyper-threading x86_64.
> > 
> > Command line to test: #hackbench 100 process 2000
> > 
> > With 2.6.26-rc1, it takes 30 seconds. But with 2.6.26-rc2/rc3, it takes 40 seconds.
> > 
> > Bisect located below patch:
> > 46151122e0a2e80e5a6b2889f595e371fe2b600d is first bad commit
> > commit 46151122e0a2e80e5a6b2889f595e371fe2b600d
> > Author: Mike Galbraith <efault@gmx.de>
> > Date:   Thu May 8 17:00:42 2008 +0200
> > 
> >     sched: fix weight calculations
> >     
> >     The conversion between virtual and real time is as follows:
> >     
> >       dvt = rw/w * dt <=> dt = w/rw * dvt
> >     
> >     Since we want the fair sleeper granularity to be in real time, we actually
> >     need to do:
> >     
> >       dvt = - rw/w * l
> > 
> > 
> > 
> > The bisect steps look stable.
> > 
> > On my core architecure machines(stoakley and tigerton), I do see improvement instead of regression,
> > like result from 31 seconds down to 28 seconds.
> > 
> > I'm not sure if hyper-threading need more cares in this situation.
> 
> Oh joy.  I'll update my poor old P4 and see what I can duplicate this.
> 
> Do you still have group scheduling enabled?
Yes.
CONFIG_GROUP_SCHED=y
CONFIG_FAIR_GROUP_SCHED=y
# CONFIG_RT_GROUP_SCHED is not set
CONFIG_USER_SCHED=y
# CONFIG_CGROUP_SCHED is not set

>   If so, can you turn it off
> and try again?  (when in doubt, grasp at any straw within reach;)
I reran the test for dozeons of times.
1) Background processes have impact on the result and cause result to fluctuate with 8~9 seconds;
2) After turning off most services (background processes), the result looks stable;
3) I tested both 2.6.26-rc1 and 2.6.26-rc2 with CONFIG_GROUP_SCHED=n. the first one's result
is about 30 seconds and the second one's result is about 31 seconds. So CONFIG_GROUP_SCHED is the key.

The previous results wree got with turning off most background processes.

I'm busy in other issues.

-yanmin

Re: hackbench regression with 2.6.26-rc2 on tulsa machine

[Mike Galbraith]

On Wed, 2008-05-21 at 12:54 +0800, Zhang, Yanmin wrote:
> On Tue, 2008-05-20 at 11:22 +0200, Mike Galbraith wrote:

> > Do you still have group scheduling enabled?
> Yes.
> CONFIG_GROUP_SCHED=y
> CONFIG_FAIR_GROUP_SCHED=y
> # CONFIG_RT_GROUP_SCHED is not set
> CONFIG_USER_SCHED=y
> # CONFIG_CGROUP_SCHED is not set
> 
> >   If so, can you turn it off
> > and try again?  (when in doubt, grasp at any straw within reach;)
> I reran the test for dozeons of times.
> 1) Background processes have impact on the result and cause result to fluctuate with 8~9 seconds;
> 2) After turning off most services (background processes), the result looks stable;
> 3) I tested both 2.6.26-rc1 and 2.6.26-rc2 with CONFIG_GROUP_SCHED=n. the first one's result
> is about 30 seconds and the second one's result is about 31 seconds. So CONFIG_GROUP_SCHED is the key.

Thanks a bunch for re-testing.

	-Mike