Re: [PerfSuite-users] Problems about using Perfsuite to monitor OpenMP program (NPB-3.2.1)

[<rk...@il...>]

Jie,

Thanks for reporting on your further experiments.  With the issue still present when using PAPI, it seems similar to an issue we have seen on our Altix.  Unfortunately, this remains an unresolved issue that may be related to the way that psrun operates internally.  I'm afraid I do not have a solution at present, but I have found that using the PerfSuite API directly produces the proper results.  If you are able and willing to do so, using the API involves inserting a call to the following routines:

call psf_hwpc_init() - from the main thread
call psf_hwpc_start() - from within a parallel region
call psf_hwpc_stop(filename) - from within a parallel region

There is an example (in C) in the PerfSuite distribution of calling the API from an OpenMP program.  You will find it in:

$PREFIX/share/perfsuite/examples/cpi/cpi-omp.c

I'm afraid I do not have a better solution at this time, but it is an issue we are following up on.

Rick


----- Original Message -----
From: "Robbie" <jj...@nu...>
To: "Rick Kufrin" <rk...@il...>
Cc: per...@li...
Sent: Wednesday, August 26, 2009 7:27:05 AM GMT -06:00 US/Canada Central
Subject: Re: [PerfSuite-users] Problems about using Perfsuite to monitor OpenMP program (NPB-3.2.1)

Rick,

Thanks for your suggestion.
However, when I tried to monitor an OpenMP program with the default
configuration file, I still got some errors and the data files are not
created as expected.

Followings are the platform information and how I measure the NPB-OMP
program with psrun.
Note here the target OpenMP program is compiled by gcc/gfortran with
-fopenmp option.


jiejiang@UT43:~/NPB3.2.1/NPB3.2-OMP/bin$ uname -a
Linux UT43 2.6.27-perfctr #2 SMP Tue Apr 28 20:29:12 CST 2009 i686
GNU/Linux
jiejiang@UT43:~/NPB3.2.1/NPB3.2-OMP/bin$ perfex -i
PerfCtr Info:
abi_version		0x05020501
driver_version		2.6.37 DEBUG
cpu_type		14 (Intel Pentium M)
cpu_features		0x7 (rdpmc,rdtsc,pcint)
cpu_khz			798049
tsc_to_cpu_mult		1
cpu_nrctrs		2
cpus			[0], total: 1
cpus_forbidden		[], total: 0

jiejiang@UT43:~/NPB3.2.1/NPB3.2-OMP/bin$ ls
bt.A  ep.A  is.A  is.B

jiejiang@UT43:~/NPB3.2.1/NPB3.2-OMP/bin$ export OMP_NUM_THREADS=2
jiejiang@UT43:~/NPB3.2.1/NPB3.2-OMP/bin$ psrun -p ./is.A


 NAS Parallel Benchmarks (NPB3.2-OMP) - IS Benchmark

 Size:  8388608  (class A)
 Iterations:  10
 Number of available threads:  2


   iteration
        1
        2
        3
        4
        5
        6
        7
        8
        9
        10


 IS Benchmark Completed
 Class           =                        A
 Size            =                  8388608
 Iterations      =                       10
 Time in seconds =                     1.46
 Total threads   =                        2
 Avail threads   =                        2
 Mop/s total     =                    57.29
 Mop/s/thread    =                    28.64
 Operation type  =              keys ranked
 Verification    =               SUCCESSFUL
 Version         =                    3.2.1
 Compile date    =              25 Aug 2009

 Compile options:
    CC           = gcc
    CLINK        = $(CC)
    C_LIB        = -lm
    C_INC        = (none)
    CFLAGS       = -O -g -fopenmp
    CLINKFLAGS   = -O -fopenmp


 Please send all errors/feedbacks to:

 NPB Development Team
 np...@na...

Inconsistency detected by ld.so: dl-close.c: 719: _dl_close: Assertion
`map->l_init_called' failed!


jiejiang@UT43:~/NPB3.2.1/NPB3.2-OMP/bin$ ls
bt.A  ep.A  is.A  is.A.0.30613.UT43.xml  is.B


The program execution finishes with the error message in the last line
and there is only ONE xml output file, not two as expected.
This also happens to papi_profile_cycles.xml configuration file.

What's wrong?

Regards,
Jie Jiang



 Rick Kufrin wrote:
> Jie,
> 
> My guess is that what is happening here is related to the use of the "itimer.xml" configuration file.  The problem is that signal delivery is not defined with POSIX threads, and the results are unpredictable.  POSIX threads enter the picture when you are using OpenMP.
> 
> Does your system happen to have kernel support for hardware counters?  If so, you may have better luck by profiling with performance counters such as total cycles rather than itimers.
> 
> Rick
> 




------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day 
trial. Simplify your report design, integration and deployment - and focus on 
what you do best, core application coding. Discover what's new with 
Crystal Reports now.  http://p.sf.net/sfu/bobj-july
_______________________________________________
PerfSuite-users mailing list
Per...@li...
https://lists.sourceforge.net/lists/listinfo/perfsuite-users