AIX perfomance

2003-09-12 Thread Rafael Garcia-Suarez
I've a mod_perl application we've developed on Linux and that
we're porting on AIX (4.3.3 and 5.2.0). The AIX boxes are
supposed to be more powerful than their Linux equivalents,
however the application is strangely slow on AIX -- the httpd
configuration being similar. And that's mod_perl 1.28.

So I'm asking for the common wisdom about performance
issues on AIX. Currently the perl I use is built with
gcc and default settings. Should I set -Dusemymalloc=y ?
Should I use the xlC or vac compilers ? Should I port
everything to mod_perl 2 ? (which I haven't succeeded
to build on AIX 4 by now BTW -- but I'm working on it.)


Re: AIX perfomance

2003-09-12 Thread Ged Haywood
Hi there,

On Fri, 12 Sep 2003, Rafael Garcia-Suarez wrote:

> we're porting on AIX (4.3.3 and 5.2.0). The AIX boxes are
> supposed to be more powerful than their Linux equivalents,
> however the application is strangely slow on AIX

You don't give much to go on.  Are they really more powerful?

What does 'powerful' mean anyway?  What discs do you have and what
interfaces do they use, how much memory, what processors, speeds, how
many mod_perl processes, how big are they, are you getting into swap,
etc...?

Have you benchmarked some simple things on the boxes?

> So I'm asking for the common wisdom about performance issues on AIX.

I don't know anything worth writing about AIX but I'd look a little
deeper into what you're doing before you start blaming the OS.

> Currently the perl I use is built with gcc and default
> settings. Should I set -Dusemymalloc=y ?  Should I use the xlC or
> vac compilers ? Should I port everything to mod_perl 2 ?

To all those questions at this stage, my answer would be 'I doubt it'.
Find out about your systems first.  There are lots of tools to help
you do that.  Start by checking the relevant sections of the Guide for
more information about performance and benchmarking.  (Or look at the
little disc activity light. :)

73,
Ged.



Re: AIX perfomance

2003-09-12 Thread William McCabe
On 9/12/03 at 2:07 PM, [EMAIL PROTECTED] (Rafael Garcia-Suarez)
wrote:

> I've a mod_perl application we've developed on Linux and that
> we're porting on AIX (4.3.3 and 5.2.0). The AIX boxes are
> supposed to be more powerful than their Linux equivalents,
> however the application is strangely slow on AIX -- the httpd
> configuration being similar. And that's mod_perl 1.28.
> 
> So I'm asking for the common wisdom about performance
> issues on AIX. Currently the perl I use is built with
> gcc and default settings. Should I set -Dusemymalloc=y ?
> Should I use the xlC or vac compilers ? Should I port
> everything to mod_perl 2 ? (which I haven't succeeded
> to build on AIX 4 by now BTW -- but I'm working on it.)

I've built many mod_perl applications on linux and moved them to AIX 4.3.3 with
no detriment at all. What do you mean by strangely slow?

BIll


Re: AIX perfomance

2003-09-12 Thread Rafael Garcia-Suarez
Ged Haywood wrote:
> Hi there,
> 
> On Fri, 12 Sep 2003, Rafael Garcia-Suarez wrote:
> 
> > we're porting on AIX (4.3.3 and 5.2.0). The AIX boxes are
> > supposed to be more powerful than their Linux equivalents,
> > however the application is strangely slow on AIX
> 
> You don't give much to go on.  Are they really more powerful?
> 
> What does 'powerful' mean anyway?  What discs do you have and what
> interfaces do they use, how much memory, what processors, speeds, how
> many mod_perl processes, how big are they, are you getting into swap,
> etc...?

Well, it's difficult to compare very different hardware, but basically
the AIX boxen have SCSI discs, more memory, etc. and they're a lot
more expensive ;-)

> Have you benchmarked some simple things on the boxes?

Benchmarking simple CPU-intensive perl scripts shows that they
tend to be consistently slower in user time on AIX.

Moreover if I survey CPU/memory usage on Linux and AIX (resp. with
top and vmstat / w) I see that the application doesn't swap memory
and that the load averages remains < 0.10.

> > So I'm asking for the common wisdom about performance issues on AIX.
> 
> I don't know anything worth writing about AIX but I'd look a little
> deeper into what you're doing before you start blaming the OS.
> 
> > Currently the perl I use is built with gcc and default
> > settings. Should I set -Dusemymalloc=y ?  Should I use the xlC or
> > vac compilers ? Should I port everything to mod_perl 2 ?
> 
> To all those questions at this stage, my answer would be 'I doubt it'.
> Find out about your systems first.  There are lots of tools to help
> you do that.  Start by checking the relevant sections of the Guide for
> more information about performance and benchmarking.  (Or look at the
> little disc activity light. :)

Thanks, I'll dig deeper.


Re: AIX perfomance

2003-09-12 Thread Ged Haywood
Hi there,

On Fri, 12 Sep 2003, Rafael Garcia-Suarez wrote:

> Benchmarking simple CPU-intensive perl scripts shows that they
> tend to be consistently slower in user time on AIX.

Assuming that the boxes aren't otherwise heavily loaded, I wonder
about the options used to compile your Perl.  For x86 architecture,
things like -mcpu=i386 will make a binary that you could run on a
steam engine but it won't be able to take advantage of the richer
instruction set on newer processors.  I don't have a great deal of
experience with other modern processors, but from the gcc 3.2.3
documentation:

"GCC defaults to `-maix32'"

and there's a '-maix64' that may be worth a look, along with the rest
of the section ('info gcc' if you have it).

Optimisation may also be an issue, but use caution.  Many packages
(e.g. the Linux kernel sources :) warn against anything more than
using -O2 with gcc for example.

73,
Ged.



Re: AIX perfomance

2003-09-12 Thread William McCabe
On 9/12/03 at 2:54 PM, [EMAIL PROTECTED] (Ged Haywood) wrote:

> > Benchmarking simple CPU-intensive perl scripts shows that they
> > tend to be consistently slower in user time on AIX.
> 
> Assuming that the boxes aren't otherwise heavily loaded, I wonder
> about the options used to compile your Perl.  For x86 architecture,
> things like -mcpu=i386 will make a binary that you could run on a
> steam engine but it won't be able to take advantage of the richer
> instruction set on newer processors.  I don't have a great deal of
> experience with other modern processors, but from the gcc 3.2.3
> documentation:
> 
> "GCC defaults to `-maix32'"
> 
> and there's a '-maix64' that may be worth a look, along with the rest
> of the section ('info gcc' if you have it).
> 
> Optimisation may also be an issue, but use caution.  Many packages
> (e.g. the Linux kernel sources :) warn against anything more than
> using -O2 with gcc for example.

I think it's pretty useless to speculate as to causes until he clarifies what
"strangely slow" means and what AIX and linux hardware he's comparing. I've got
a lot of experience with mod_perl on both linux and AIX and can state
categorically that there are no typical conditions which would cause AIX run
"strangely slowly" compared to linux on comparable hardware.

Bill


Re: AIX perfomance

2003-09-12 Thread Ged Haywood
Hi there,

On Fri, 12 Sep 2003, William McCabe wrote:

> I've got a lot of experience with mod_perl on both linux and AIX and
> can state categorically that there are no typical conditions which
> would cause AIX run "strangely slowly" compared to linux on
> comparable hardware.

That's useful information, thanks Bill.  We don't see much about AIX
here but it seems to me that it's been becoming more common recently.

Roughly what hardware setups do you generally work with, and what
differences are notable between Linux and AIX when running mod_perl
servers?  (If that's not too long a piece of string to measure:).
Are there situations where you'd prefer one or the other, if so why?

73,
Ged.



Re: AIX perfomance

2003-09-12 Thread Perrin Harkins
On Fri, 2003-09-12 at 09:35, Rafael Garcia-Suarez wrote:
> Benchmarking simple CPU-intensive perl scripts shows that they
> tend to be consistently slower in user time on AIX.

Are these mod_perl scripts or just Perl?  If you benchmark some simple
Perl scripts that don't run under mod_perl and they show the same trend,
then you can eliminate mod_perl as a possible source of the problem.

You should also verify that your versions of Perl, apache, and mod_perl
are exactly the same on both systems.

- Perrin


Re: AIX perfomance

2003-09-15 Thread Patrick Mulvany
On Fri, Sep 12, 2003 at 04:12:00PM +0100, Ged Haywood wrote:
> Hi there,
> 
> On Fri, 12 Sep 2003, William McCabe wrote:
> 
> > I've got a lot of experience with mod_perl on both linux and AIX and
> > can state categorically that there are no typical conditions which
> > would cause AIX run "strangely slowly" compared to linux on
> > comparable hardware.
> 
> That's useful information, thanks Bill.  We don't see much about AIX
> here but it seems to me that it's been becoming more common recently.
> 

AIX will of course have an massive growth in popularity now that SCO has withdrwan 
IBMs un*x licence. ;*>

Paddy




Re: AIX perfomance

2003-09-17 Thread William McCabe
Hi Ged,

On 9/12/03 at 4:12 PM, [EMAIL PROTECTED] (Ged Haywood) wrote:

> 
> Roughly what hardware setups do you generally work with, and what
> differences are notable between Linux and AIX when running mod_perl
> servers?  (If that's not too long a piece of string to measure:).
> Are there situations where you'd prefer one or the other, if so why?

Sorry for the slow response; I've been out of town. Most of my mod_perl/AIX
systems are used to generate organizational performance reports, basically
data-mart type stuff, which is very DBI (DB2) and computationally-intensive, and
also often invlove running COBOL binaries which have been ported from OS/390 and
run via RPC::XML. If the need to run COBOL is absent from a project, then I
usually deploy on Lintel, since procurement is so much easier. I never rely on
OS ditributions of perl, apache, or mod_perl so my working enviroment is always
identical. As I mentioned earlier, Rafael should not be experience slowness on
AIX unless he's comparing dated RS/6000 hardware with new Intel. Scalability,
especially with big SMP iron, still favors the RS/6000 though at a colossal
cost.

Bill