FC4 test (x86_64) poor performance

[Vlad]

Hello,

I've ran a series of tests on FC4 x86_64 and FC3 x86_64. Found that
the performance is quite poor under FC4 (Athlon64 on nForce 4 Ultra).

I've compiled apache+php and measured max requests per second (rps) of
a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
the apache's 'ab' application. The maximum RPS I managed for FC4 is
440 while FC3 showed up to 730. Network is not the issue cause I did
measurement via lo0.

It's the same server and I've ran yum update before tests in both
cases. Any ideas?

[Patrick W. Barnes]

Vlad wrote:
> Hello,
>
> I've ran a series of tests on FC4 x86_64 and FC3 x86_64. Found that
> the performance is quite poor under FC4 (Athlon64 on nForce 4 Ultra).

I've had no performance problems with x86_64 installations, on any of
the chipsets I've tried.  In some cases, the speed has improved
dramatically.
> I've compiled apache+php and measured max requests per second (rps) of
> a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
> the apache's 'ab' application. The maximum RPS I managed for FC4 is
> 440 while FC3 showed up to 730.

Perhaps you should try a different benchmark without recompiling
software.  Version or configuration differences could easily be causing
your problems.  Did you at least use the SRPMs?
> Network is not the issue cause I did
> measurement via lo0.
>
> It's the same server and I've ran yum update before tests in both
> cases. Any ideas?

Try other benchmarks and see if you can narrow down the problem a bit,
or determine if there even is a problem to begin with.

--
Patrick "The N-Man" Barnes
[e-mail]

www.n-man.com
--

[Dave Jones]

On Sun, May 29, 2005 at 09:32:56AM -0400, Vlad wrote:
> Hello,
>
> I've ran a series of tests on FC4 x86_64 and FC3 x86_64. Found that
> the performance is quite poor under FC4 (Athlon64 on nForce 4 Ultra).
>
> I've compiled apache+php and measured max requests per second (rps) of
> a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
> the apache's 'ab' application. The maximum RPS I managed for FC4 is
> 440 while FC3 showed up to 730. Network is not the issue cause I did
> measurement via lo0.
>
> It's the same server and I've ran yum update before tests in both
> cases. Any ideas?

If you are running a kernel earlier than revision 2.6.11-1.1355, it'll
have extra debugging enabled which will impact memory allocations.

Are you running the latest rawhide ?

Dave

[Sean Bruno]

On Sun, 2005-05-29 at 18:11 -0400, Dave Jones wrote:
> On Sun, May 29, 2005 at 09:32:56AM -0400, Vlad wrote:
> > Hello,
> >
> > I've ran a series of tests on FC4 x86_64 and FC3 x86_64. Found that
> > the performance is quite poor under FC4 (Athlon64 on nForce 4 Ultra).
> >
> > I've compiled apache+php and measured max requests per second (rps) of
> > a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
> > the apache's 'ab' application. The maximum RPS I managed for FC4 is
> > 440 while FC3 showed up to 730. Network is not the issue cause I did
> > measurement via lo0.
> >
> > It's the same server and I've ran yum update before tests in both
> > cases. Any ideas?
>
> If you are running a kernel earlier than revision 2.6.11-1.1355, it'll
> have extra debugging enabled which will impact memory allocations.
>
> Are you running the latest rawhide ?
> Dave

Dave:

Should those of us running off of rawhide add anything extra to our bug
reports due to this extra debugging?

Sean

[Dave Jones]

On Sun, May 29, 2005 at 03:37:12PM -0700, Sean Bruno wrote:
> On Sun, 2005-05-29 at 18:11 -0400, Dave Jones wrote:
> > On Sun, May 29, 2005 at 09:32:56AM -0400, Vlad wrote:
> > > Hello,
> > >
> > > I've ran a series of tests on FC4 x86_64 and FC3 x86_64. Found that
> > > the performance is quite poor under FC4 (Athlon64 on nForce 4 Ultra).
> > >
> > > I've compiled apache+php and measured max requests per second (rps) of
> > > a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
> > > the apache's 'ab' application. The maximum RPS I managed for FC4 is
> > > 440 while FC3 showed up to 730. Network is not the issue cause I did
> > > measurement via lo0.
> > >
> > > It's the same server and I've ran yum update before tests in both
> > > cases. Any ideas?
> >
> > If you are running a kernel earlier than revision 2.6.11-1.1355, it'll
> > have extra debugging enabled which will impact memory allocations.
> >
> > Are you running the latest rawhide ?
>
> Dave:
>
> Should those of us running off of rawhide add anything extra to our bug
> reports due to this extra debugging?

Either I've misinterpreted your question, or you misinterpreted my
previous mail.  The current kernels have the debugging disabled.

Dave

[Sean Bruno]

On Sun, 2005-05-29 at 18:44 -0400, Dave Jones wrote:
> On Sun, May 29, 2005 at 03:37:12PM -0700, Sean Bruno wrote:
> > On Sun, 2005-05-29 at 18:11 -0400, Dave Jones wrote:
> > > On Sun, May 29, 2005 at 09:32:56AM -0400, Vlad wrote:
> > > > Hello,
> > > >
> > > > I've ran a series of tests on FC4 x86_64 and FC3 x86_64. Found that
> > > > the performance is quite poor under FC4 (Athlon64 on nForce 4 Ultra).
> > > >
> > > > I've compiled apache+php and measured max requests per second (rps) of
> > > > a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
> > > > the apache's 'ab' application. The maximum RPS I managed for FC4 is
> > > > 440 while FC3 showed up to 730. Network is not the issue cause I did
> > > > measurement via lo0.
> > > >
> > > > It's the same server and I've ran yum update before tests in both
> > > > cases. Any ideas?
> > >
> > > If you are running a kernel earlier than revision 2.6.11-1.1355, it'll
> > > have extra debugging enabled which will impact memory allocations.
> > >
> > > Are you running the latest rawhide ?
> >
> > Dave:
> >
> > Should those of us running off of rawhide add anything extra to our bug
> > reports due to this extra debugging?
>
> Either I've misinterpreted your question, or you misinterpreted my
> previous mail.  The current kernels have the debugging disabled.
> Dave

Backwards reading doing was I.

Never mind.

Sean

[Vlad]

Dave,

I did the tests several days ago; following your advise I just ran yum
update and tested once again with 2.6.11-1.1363_FC4. Things are much
improved now.

thank you.

On 5/29/05, Dave Jones  wrote:
> On Sun, May 29, 2005 at 09:32:56AM -0400, Vlad wrote:
> > Hello,
> >
> > I've ran a series of tests on FC4 x86_64 and FC3 x86_64. Found that
> > the performance is quite poor under FC4 (Athlon64 on nForce 4 Ultra).
> >
> > I've compiled apache+php and measured max requests per second (rps) of
> > a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
> > the apache's 'ab' application. The maximum RPS I managed for FC4 is
> > 440 while FC3 showed up to 730. Network is not the issue cause I did
> > measurement via lo0.
> >
> > It's the same server and I've ran yum update before tests in both
> > cases. Any ideas?
>
> If you are running a kernel earlier than revision 2.6.11-1.1355, it'll
> have extra debugging enabled which will impact memory allocations.
>
> Are you running the latest rawhide ?
>             Dave

[Dave Jones]

On Sun, May 29, 2005 at 06:40:55PM -0400, Vlad wrote:
> > > I've compiled apache+php and measured max requests per second (rps) of
> > > a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
> > > the apache's 'ab' application. The maximum RPS I managed for FC4 is
> > > 440 while FC3 showed up to 730. Network is not the issue cause I did
> > > measurement via lo0.
> > >
> > > It's the same server and I've ran yum update before tests in both
> > > cases. Any ideas?
> >
> > If you are running a kernel earlier than revision 2.6.11-1.1355, it'll
> > have extra debugging enabled which will impact memory allocations.
>
> I did the tests several days ago; following your advise I just ran yum
> update and tested once again with 2.6.11-1.1363_FC4. Things are much
> improved now.

Great! However, don't tease :-)  I'm sure I'm not the only person
curious what RPS you now achieve :-)

Dave

[Vlad]

ok, here is the complete details:

Athlon64 3000+ (it's +20Mhz overclocked on the bus, i.e. running on
220Mhz x 9 - roughly on par with Athlon64 3200+).
NForce Ultra (MB: tyan k8e) with 1Gig.

php 5.0.4  + 1.3.33 statically compiled with -O3 -msse3 -m3dnow
-march=athlon64 -mcpu=athlon64; (can't recall if I did unroll loops
and omit frame pointer)

no extra switches or additional modules for configure.

the test.php script looked like

<?php
phpinfo();
phpinfo(INFO_MODULES);
?>

then tested with 
/usr/local/apache/bin/ab -c 50 -n 20000 http://localhost/test.php

FC4 x86_64, kernel is 2.6.11-1.1363_FC4 - the RPS is 730.
FC3 x86_64, latests kernel - same results.
FC3 i386, latests kernel - 570 RPS.
FreeBSD 5.4-RELEASE (x86_64, stock kernel) - 670 RPS

same hardware in all tests, same optimization switches during compilation.

On 5/29/05, Dave Jones  wrote:
> On Sun, May 29, 2005 at 06:40:55PM -0400, Vlad wrote:
> > > > I've compiled apache+php and measured max requests per second (rps) of
> > > > a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
> > > > the apache's 'ab' application. The maximum RPS I managed for FC4 is
> > > > 440 while FC3 showed up to 730. Network is not the issue cause I did
> > > > measurement via lo0.
> > > >
> > > > It's the same server and I've ran yum update before tests in both
> > > > cases. Any ideas?
> > >
> > > If you are running a kernel earlier than revision 2.6.11-1.1355, it'll
> > > have extra debugging enabled which will impact memory allocations.
> >
> > I did the tests several days ago; following your advise I just ran yum
> > update and tested once again with 2.6.11-1.1363_FC4. Things are much
> > improved now.
>
> Great! However, don't tease :-)  I'm sure I'm not the only person
> curious what RPS you now achieve :-)
>             Dave

[Vlad]

I forgot to mention -m64 flag that I also used.

On 5/29/05, Vlad  wrote:
> ok, here is the complete details:
>
> Athlon64 3000+ (it's +20Mhz overclocked on the bus, i.e. running on
> 220Mhz x 9 - roughly on par with Athlon64 3200+).
> NForce Ultra (MB: tyan k8e) with 1Gig.
>
> php 5.0.4  + 1.3.33 statically compiled with -O3 -msse3 -m3dnow
> -march=athlon64 -mcpu=athlon64; (can't recall if I did unroll loops
> and omit frame pointer)
>
> no extra switches or additional modules for configure.
>
> the test.php script looked like
>
> <?php
> phpinfo();
> phpinfo(INFO_MODULES);
> ?>
>
> then tested with
> /usr/local/apache/bin/ab -c 50 -n 20000 http://localhost/test.php
>
> FC4 x86_64, kernel is 2.6.11-1.1363_FC4 - the RPS is 730.
> FC3 x86_64, latests kernel - same results.
> FC3 i386, latests kernel - 570 RPS.
> FreeBSD 5.4-RELEASE (x86_64, stock kernel) - 670 RPS
>
> same hardware in all tests, same optimization switches during compilation.
>
> On 5/29/05, Dave Jones  wrote:
> > On Sun, May 29, 2005 at 06:40:55PM -0400, Vlad wrote:
> > > > > I've compiled apache+php and measured max requests per second (rps) of
> > > > > a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
> > > > > the apache's 'ab' application. The maximum RPS I managed for FC4 is
> > > > > 440 while FC3 showed up to 730. Network is not the issue cause I did
> > > > > measurement via lo0.
> > > > >
> > > > > It's the same server and I've ran yum update before tests in both
> > > > > cases. Any ideas?
> > > >
> > > > If you are running a kernel earlier than revision 2.6.11-1.1355, it'll
> > > > have extra debugging enabled which will impact memory allocations.
> > >
> > > I did the tests several days ago; following your advise I just ran yum
> > > update and tested once again with 2.6.11-1.1363_FC4. Things are much
> > > improved now.
> >
> > Great! However, don't tease :-)  I'm sure I'm not the only person
> > curious what RPS you now achieve :-)
> >             Dave
>
> --
>
> Vlad

[Vlad]

lastly, if it will make someones choice of the platform easier:
same test.php on P4-3Ghz-1Gb, FC3-386 shows 540 RPS - comparable to
Athlon 3000+ results on FC3 i386, but noticeably lesser than what same
Athlon shows in 64 environment.

On 5/29/05, Vlad  wrote:
> I forgot to mention -m64 flag that I also used.
>
> On 5/29/05, Vlad  wrote:
> > ok, here is the complete details:
> >
> > Athlon64 3000+ (it's +20Mhz overclocked on the bus, i.e. running on
> > 220Mhz x 9 - roughly on par with Athlon64 3200+).
> > NForce Ultra (MB: tyan k8e) with 1Gig.
> >
> > php 5.0.4  + 1.3.33 statically compiled with -O3 -msse3 -m3dnow
> > -march=athlon64 -mcpu=athlon64; (can't recall if I did unroll loops
> > and omit frame pointer)
> >
> > no extra switches or additional modules for configure.
> >
> > the test.php script looked like
> >
> > <?php
> > phpinfo();
> > phpinfo(INFO_MODULES);
> > ?>
> >
> > then tested with
> > /usr/local/apache/bin/ab -c 50 -n 20000 http://localhost/test.php
> >
> > FC4 x86_64, kernel is 2.6.11-1.1363_FC4 - the RPS is 730.
> > FC3 x86_64, latests kernel - same results.
> > FC3 i386, latests kernel - 570 RPS.
> > FreeBSD 5.4-RELEASE (x86_64, stock kernel) - 670 RPS
> >
> > same hardware in all tests, same optimization switches during compilation.
> >
> > On 5/29/05, Dave Jones  wrote:
> > > On Sun, May 29, 2005 at 06:40:55PM -0400, Vlad wrote:
> > > > > > I've compiled apache+php and measured max requests per second (rps) of
> > > > > > a simple php script (output of phpinfo(); php_info(INFO_MODULES)) with
> > > > > > the apache's 'ab' application. The maximum RPS I managed for FC4 is
> > > > > > 440 while FC3 showed up to 730. Network is not the issue cause I did
> > > > > > measurement via lo0.
> > > > > >
> > > > > > It's the same server and I've ran yum update before tests in both
> > > > > > cases. Any ideas?
> > > > >
> > > > > If you are running a kernel earlier than revision 2.6.11-1.1355, it'll
> > > > > have extra debugging enabled which will impact memory allocations.
> > > >
> > > > I did the tests several days ago; following your advise I just ran yum
> > > > update and tested once again with 2.6.11-1.1363_FC4. Things are much
> > > > improved now.
> > >
> > > Great! However, don't tease :-)  I'm sure I'm not the only person
> > > curious what RPS you now achieve :-)
> > >             Dave
> >
> > --
> >
> > Vlad
>
> --
>
> Vlad

[Dave Jones]

On Sun, May 29, 2005 at 11:41:24PM -0400, Vlad wrote:
> FC4 x86_64, kernel is 2.6.11-1.1363_FC4 - the RPS is 730.
> FC3 x86_64, latests kernel - same results.
> FC3 i386, latests kernel - 570 RPS.
> FreeBSD 5.4-RELEASE (x86_64, stock kernel) - 670 RPS

I'd suspect the reason FC3 sucks right now on this benchmark is that
2.6.11 contained the 4-level page table support for AMD64, though
it wasn't optimised until 2.6.12rc.

At least its the only thing that comes to mind that could explain
such a huge delta between the two. Good to know things are back
to normal however.

Dave

[Vlad]

wait, you mean the i386 version of FC3? case FC3 x86_64 shows same
numbers as FC4 x86_64 for me....

On 5/29/05, Dave Jones  wrote:
> On Sun, May 29, 2005 at 11:41:24PM -0400, Vlad wrote:
> > FC4 x86_64, kernel is 2.6.11-1.1363_FC4 - the RPS is 730.
> > FC3 x86_64, latests kernel - same results.
> > FC3 i386, latests kernel - 570 RPS.
> > FreeBSD 5.4-RELEASE (x86_64, stock kernel) - 670 RPS
>
> I'd suspect the reason FC3 sucks right now on this benchmark is that
> 2.6.11 contained the 4-level page table support for AMD64, though
> it wasn't optimised until 2.6.12rc.
>
> At least its the only thing that comes to mind that could explain
> such a huge delta between the two. Good to know things are back
> to normal however.
>             Dave

[Dave Jones]

On Mon, May 30, 2005 at 12:03:44AM -0400, Vlad wrote:
> > > FC4 x86_64, kernel is 2.6.11-1.1363_FC4 - the RPS is 730.
> > > FC3 x86_64, latests kernel - same results.
> > > FC3 i386, latests kernel - 570 RPS.
> > > FreeBSD 5.4-RELEASE (x86_64, stock kernel) - 670 RPS
> >
> > I'd suspect the reason FC3 sucks right now on this benchmark is that
> > 2.6.11 contained the 4-level page table support for AMD64, though
> > it wasn't optimised until 2.6.12rc.
> >
> > At least its the only thing that comes to mind that could explain
> > such a huge delta between the two. Good to know things are back
> > to normal however.
>
> wait, you mean the i386 version of FC3? case FC3 x86_64 shows same
> numbers as FC4 x86_64 for me....

Oh, in that case, I have no explanation, as the 4-level pagetables
should have only affected x86-64.

Dave