perfsuite-users — Mailing list for users of PerfSuite software

[PerfSuite-users] Re: PerfSuite problem - multiplexing not working?

[<rk...@nc...>]

p.s. Kursad, there is something else that I thought might be useful
to mention: even though the Pentium 4 has 18 counters, multiplexing
may still be enabled even when counting less than 18 events.  That's
because the counters are not symmetric - they cannot all count the
same events.  Additionally, some events require two counters (for
example, retired floating point instructions), so that takes up
additional resources.   If you get an output XML document, you can
check whether multiplexing was enabled either directly (it's listed
as one of the element attributes) or in the psprocess report.

Rick

[PerfSuite-users] Re: PerfSuite problem - multiplexing not working?

[<rk...@nc...>]

Kursad,

Sorry to hear about the problems you're having but maybe we can
figure it out.  We have a similar setup at NCSA both on PIII and Xeon
systems and are using both PAPI 2 and 3 without problems.  As you
point out, cannot use PAPI 2 on the Pentium 4 system though.
It sounds to me like your PAPI install is fine.

I'll copy the SourceForge mailing list in case this might be of
interest to others.

There are a couple things I can think of to suggest:  first, are you
running psrun with a command that consumes a reasonable amount
of CPU time?  If not, you might get all zeroes due to the nature
of multiplexing (the counters may not have had sufficient time
to switch to actually collect counts).

If you are, then the next step is to configure PerfSuite with debugging
enabled (--enable-debug) and recompile.  That will cause a lot of tracing 
output to come (not what you'd want to use in production, but helpful
in seeing where things are getting out of line.  You'll also want
to set the environment variable PS_DEBUG to the value 3 or higher.

Also, you might try a run outside of psrun.  This is pretty simple
to do directly through the API.  You'll want to add the following
function calls (assuming you're using C/C++):

ps_hwpc_init();
ps_hwpc_start();
... do some computation here
ps_hwpc_stop("test");

One thing I should point out about the PerfSuite test suite is that
it has to be flexible enough to work regardless of the presence of
PAPI, so it's not as thorough as it could be.  That means the tests
of psrun may succeed when in fact the install is not quite right.

Let me know if any of these suggestions help - would like to get
you set up properly.

Rick

p.s. on the PIII with its two counters, what that means is that you
cannot use any "derived" events when multiplexing, because one counter
is always taken up for cycles, leaving only one left for other uses.

> ============ Kursad Albayraktaroglu's message of Jul  1,  8:41pm ================
> 
> We are trying to use Perfsuite 0.6.1b5 on a Linux/x86 machine, kernel
> version 2.4.22. We are using PAPI 3 (we're aware PAPI 3 support is alpha,
> but the unfortunate Linux kernel/GCC/PAPI combination that we have to use
> restricts us to this setup.)
> 
> All Perfsuite and PAPI tests during installation are completed
> successfully. When we run psrun with a configuration file that lists any
> of the two supported PAPI events, accurate event counts are reported.
> When more than two PAPI events are selected, we get zeros for all of
> them.
> 
> Figuring that this might have something to do with the 2 counter
> limitation of the Pentium 3 CPU on our system, we built the same software
> on a P4 system that has 18 counters. Same result : When more than 2 (not
> 18!) PAPI events are selected, all counts are reported as zero.
> 
> It would seem that multiplexing is not working, but we're perplexed about
> the situation on the P4 machine. Do you have any suggestions (short of
> switching to PAPI 2 which we can't do) or ideas on what we're doing wrong?
> 
> Regards,
> 
> --
> Kursad Albayraktaroglu
> 
> Ph.D. Candidate,
> Department of Electrical & Computer Engineering,
> University of Maryland, College Park
> 
> "Where ... the ENIAC is equipped with 18,000 vacuum tubes and weighs
> 30 tons, computers in the future may have 1,000 vacuum tubes and
> perhaps weigh just 1-1/2 tons."
> 
> 	Popular Mechanics, March 1949, p.258
> ============ End of Kursad Albayraktaroglu's message of Jul  1,  8:41pm =========

[PerfSuite-users] PerfSuite 0.6.1b5 and presentations available

[<rk...@nc...>]

I thought I'd cross-posted this to both perfsuite-announce and
perfsuite-users but looking back perhaps not.  Apologies for a
repeat.

This is to let you know that PerfSuite 0.6.1 beta 5 was released
earlier this month.  This is a substantial update that includes
AMD support as well as updates to bring it in synch with the
most recent releases of PAPI 3.  The AMD additions could use
additional testing so any feedback on that would be welcome.

Also, there are a few background presentations about PerfSuite
now available at http://perfsuite.ncsa.uiuc.edu/ and the
SourceForge web site - http://perfsuite.sourceforge.net/ that
may be helpful to people getting started.

Rick

[PerfSuite-users] Re: function level profiling

[<rk...@nc...>]

Kshitiz,

For some reason, SourceForge mailing lists didn't send me a copy of this
email but I see that it made it on the archives, so here's an
answer to your post:

You were looking in the right place for sample configuration files to
customize the hardware performance measurement, but it's not required
that you specify those exact files to get non-default behavior.  The
"recipe" for doing profiling versus the default counting is to:

1. Use a configuration file with the XML root document type as
   <ps_hwpc_profile>.  This document should specify one event only,
   the event to be used for sampling.  Here's an example:

   <?xml version="1.0" encoding="UTF-8" ?>
   <ps_hwpc_profile class="PAPI">
     <ps_hwpc_event type="preset" name="PAPI_L2_TCM" threshold="100000"/>
   </ps_hwpc_profile>

   Here, the event used is PAPI_L2_TCM, sampled at a frequency of 100,000
   events.  There isn't an exact frequency I can think of to recommend
   for arbitrary events, although if you sample based on total cycles
   choosing an equivalent to 10 ms should approximate what "gprof" does.

2. Make sure your application is compiled with symbols retained by
   using the "-g" option to the compiler.

3. Run psrun with the option "-c FILE",  where FILE is replaced by
   the name of the alternate config file.  The XML output should be
   a container with the total samples taken for each program address.
   These will usually be a lot larger than the counting equivalent.

4. Run psprocess with the option "-e EXE", with EXE being the name of
   your executable.  This uses binutils to map program addresses to
   source code lines.  By default, you should get a breakdown by
   module (executable and shared libraries), file, function, and line.

Hope this helps, let me know if you still have problems.

Rick


> Hi!
>  
>  I was trying to get function-wise statistical profile data by using 
>  perfsuite. The xml configuration for profiling was not found where 
>  indicated in the online perfsuite presentation.
>  Other places I looked were in perfsuite/share/perfsuite/xml/pshwpc where 
>  there were a lot of configs. I tried psrun with most of them, but the data 
>  generated by psrun didn't have function wise information.
>  I am using perfsuite with PAPI3 on a Xeon machine running Linux 2.4, and I 
>  used -q while compiling with g++ to retain symbol information in the
>  object files.
>  
>  Which configuration file should I be using?
>  
>  Thanks!
>  -Kshitiz
>  
>  
>  http://www.crhc.uiuc.edu/~kmalik

[PerfSuite-users] function level profiling

[Kshitiz M. <km...@cr...>]

Hi!

I was trying to get function-wise statistical profile data by using 
perfsuite. The xml configuration for profiling was not found where 
indicated in the online perfsuite presentation.
Other places I looked were in perfsuite/share/perfsuite/xml/pshwpc where 
there were a lot of configs. I tried psrun with most of them, but the data 
generated by psrun didn't have function wise information.
I am using perfsuite with PAPI3 on a Xeon machine running Linux 2.4, and I 
used -q while compiling with g++ to retain symbol information in the
object files.

Which configuration file should I be using?

Thanks!
-Kshitiz


www.crhc.uiuc.edu/~kmalik

Re: [PerfSuite-users] Re: Mailman results for PerfSuite-users

[<rk...@nc...>]

Bo,

Thanks for sending both the portion that seemed confusing and
the output of the test suite.  It looks as though things are OK
from what you've sent.  The test suite for that particular routine
tries to use both valid and invalid input to check its error
handling.  You can see what it's doing if you change into
the src/libperfsuite/tests directory and run the command:

./testsuite -v 12

You should get output from three separate runs.  The first two
use valid input and the last doesn't (that's the output I think
you've seen).

Hope you are having good luck otherwise with PerfSuite,

Rick


> Hi,
> I just installed PerfSuit on Itanium 2, and
> ran the testsuit. Want to make sure what I got is
> not because of a bug or sth. I got these:
>  
> > > [boliu@mckinley testsuite.dir]$ cat at-stderr
> > > ps_addrmap failed
> > > [boliu@mckinley testsuite.dir]$ cat at-stder1
> > > ps_addrmap failed
> > > [boliu@mckinley testsuite.dir]$ cat at-stdout
> > > ==== ps_addrmap output ======== 
> > > 
> 
> But the testing result declares OK in the 12th
> item:
> 
> > > 
> > > [boliu@mckinley tests]$ ./testsuite
> > > ## ------------------------------------------- ##
> > > ## PerfSuite 0.6.1b3 test suite: Core Library. ##
> > > ## ------------------------------------------- ##
> > >   1: Hello, world!                               
> > ok
> > >   2: CPU identification                          
> > ok
> > >   3: CPU speed information                       
> > ok
> > >   4: CPU speed information (Fortran)             
> > ok
> > >   5: CPU usage                                   
> > ok
> > >   6: string handling                             
> > ok
> > >   7: error reporting                             
> > ok
> > >   8: error reporting (Fortran)                   
> > ok
> > >   9: proc filesystem                             
> > ok
> > >  10: memory usage reporting                      
> > ok
> > >  11: memory usage reporting (Fortran)            
> > ok
> > >  12: load map address inquiry                    
> > ok
> > >  
> > > ## ------------- ##
> > > ## Test results. ##
> > > ## ------------- ##
> > >  
> 
>  
> Can anyone help with this?
>  
> Thanks,
> Bo

[PerfSuite-users] Re: Mailman results for PerfSuite-users

[Emma <liu...@ya...>]

Hi,
I just installed PerfSuit on Itanium 2, and
ran the testsuit. Want to make sure what I got is
not because of a bug or sth. I got these:
 
> > [boliu@mckinley testsuite.dir]$ cat at-stderr
> > ps_addrmap failed
> > [boliu@mckinley testsuite.dir]$ cat at-stder1
> > ps_addrmap failed
> > [boliu@mckinley testsuite.dir]$ cat at-stdout
> > ==== ps_addrmap output ======== 
> > 

But the testing result declares OK in the 12th
item:

> > 
> > [boliu@mckinley tests]$ ./testsuite
> > ## ------------------------------------------- ##
> > ## PerfSuite 0.6.1b3 test suite: Core Library. ##
> > ## ------------------------------------------- ##
> >   1: Hello, world!                               
> ok
> >   2: CPU identification                          
> ok
> >   3: CPU speed information                       
> ok
> >   4: CPU speed information (Fortran)             
> ok
> >   5: CPU usage                                   
> ok
> >   6: string handling                             
> ok
> >   7: error reporting                             
> ok
> >   8: error reporting (Fortran)                   
> ok
> >   9: proc filesystem                             
> ok
> >  10: memory usage reporting                      
> ok
> >  11: memory usage reporting (Fortran)            
> ok
> >  12: load map address inquiry                    
> ok
> >  
> > ## ------------- ##
> > ## Test results. ##
> > ## ------------- ##
> >  

 
Can anyone help with this?
 
Thanks,
Bo


__________________________________
Do you Yahoo!?
Yahoo! SiteBuilder - Free web site building tool. Try it!
http://webhosting.yahoo.com/ps/sb/