Ceci est une ancienne révision du document !


TAU

TAU is an MPI program analyzer. It generates all necessary information during the program execution, so that it can be read after it and displayed by a specialized program.

  • 2.25 compiled with GCC and Open MPI
  • 2.25 compiled with ICC and Open MPI

use module avail tau to view installed versions.

example :

$ module avail perf/tau
  • Dynamic Instrumentation: No change to your application's source or binary code is necessary. Using the tau_exec command before your executable generates both profile and trace data.
  • Compiler-Based Instrumentation: TAU can also be used to compile your application using wrapper scripts for mpicc, mpicxx, and mpif90. Tau instruments the binary code without changing the source. Using the tau_exec command before the TAU-compiled executable generates profile and trace data. Data is collected on both MPI calls and your application's function calls.
To display trace files we need to enable it first export TAU_TRACE=1
Others env variables:
  • TAU_PROFILE=1 (default)
  • TAU_CALL=0 (dfault)

</code>

  1. Load the needed tau
    $ module load perf/tau/icc
  2. Compile
    $ mpicc hello.c -o hello
  3. Dynamic instrumentation
    $ mpirun -np $NSLOTS tau_exec ./hello
  4. Run TAU's recompiling tool
    $ tau_treemerge.pl
  5. After this it will have generated some output files, merge them with the next command
    $ tau2slog2 tau.trc tau.edf -o tau.slog2
  6. Now just open the new merged file with the default displaying tool
    $ jumpshot tau.slog2
  7. For profiling just call pprof and paraprof programs to view profilers results

For each compiler type there is TAU wrapper, for example:

  • tau_cxx.sh for C++ compiler
  • tau_f90.sh for Fortarn 90 compiler
  • tau_cc.sh fot C compiler

Tau warapper uses $TAU_MAKEFILE to compile programs.

  1. Compile programs
    tau_cc.sh hello.c -o hello
  2. Run program as usual
    mpirun -np $NSLOTS ./hello
  3. Follow steps 4, 5, 6 like for dynamic instrumentation

You can also view profilers information with: pprof or paraprof

Let's profile/trace program (simple Jacobi iteration)

Tracing

$ module load perf/tau/icc
 
$ mpicc jacobi.c -o jacobi
 
$ mpirun -np 4 tau_exec ./jacobi
 
$ tau_treemerge.pl
 
$ tau2slog2 tau.trc tau.edf -o tau.slog2
 
$ jumpshot tau.slog2


Profiling
We need to pack all the profiles.* files in a unique one:

$ paraprof --pack tau.ppk

Now, we can visualize the tau.ppk file by using Paraprof

$ paraprof
click to enlarge

  • outils_analyse/tau.1460498416.txt.gz
  • Dernière modification : 2016/04/12 22:00
  • de admin