[mpich-discuss] Profiling MPICH Source Code

Amer, Abdelhalim aamer at anl.gov
Fri Apr 5 12:32:54 CDT 2019


I'll throw in a couple of options that are good at large scale:

Open Source: HPCToolkit (http://hpctoolkit.org/)
Commercial: Arm MAP

Similar to Jeff's comment about the other tools, you need to build MPICH 
with debug symbols for these tools to give you useful information.

Halim
www.mcs.anl.gov/~aamer

On 4/3/19 3:00 PM, Jeff Hammond via discuss wrote:
> First, you need to compile MPICH with debug symbols. 
> https://wiki.mpich.org/mpich/index.php/Building_MPICH_For_Development has details.
> 
> Second, use a profiler like Linux perf or Intel Vtune to see what the 
> hotspots are.
> 
> Jeff
> 
> On Wed, Apr 3, 2019 at 11:54 AM SANJAY KUMAR via discuss 
> <discuss at mpich.org <mailto:discuss at mpich.org>> wrote:
> 
>     Here at Indian Institute of Technology Kanpur, Computer Science
>     Department, we are trying to profile mpich source code itself. So as to
>     analyze the internal function calls made by the mpich functions like
>     MPI_Bcast etc.  I have tried so many profilers like IPM, gprof to
>     compile the MPICH source code. But It turned out nothing works. I have
>     also so many flags but nothing worked out. Could you please tell me how
>     to profile the mpich source code itself.
> 
> 
>     Thanks
> 
>     Sanjay Kumar
> 
>     Masters of Technology
> 
>     Compute Science Department
> 
>     Indian Institute of Technology Kanpur
> 
>     _______________________________________________
>     discuss mailing list discuss at mpich.org <mailto:discuss at mpich.org>
>     To manage subscription options or unsubscribe:
>     https://lists.mpich.org/mailman/listinfo/discuss
> 
> 
> 
> -- 
> Jeff Hammond
> jeff.science at gmail.com <mailto:jeff.science at gmail.com>
> http://jeffhammond.github.io/
> 
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss
> 


More information about the discuss mailing list