Valgrind

From ScorecWiki

Contents 1 Parallel Execution 1.1 Threads 2 Mem-check 3 Massif - profile heap usage 4 Documentation

Parallel Execution

Setup your environment

module load valgrind

Run Valgrind independently on each MPI process of a parallel program using MPI named 'helloWorldMPI' for this example:

 /PATH/TO/mpirun -np <NUM PROCESSES> /PATH/TO/valgrind [valgrind options] ./helloWorldMPI

Useful options:

To output one leak-report file per-process in the xml format, which can be read by Valkyrie:

 --xml=yes --xml-file=valgrind.%p

To output one leak-report file per-process in the format which can be read by Alleyoop:

 --log-file=valgrind.%p

The following alternative file format specifier appends the MPI rank to the file name when OpenMPI is used. Note OMPI_COMM_WORLD_RANK is an environment variable set by OpenMPI.

 --log-file=valgrind.%q{OMPI_COMM_WORLD_RANK}

Threads

If your code uses threads they can be ran with equal scheduling priority via the --fair-sched=yes argument.

Note, without this argument performance will be poor.

Mem-check

To enable full leak checking:

 --leak-check=full

Massif - profile heap usage

Run with threads given equal scheduling priority:

 mpirun -np $SLURM_NPROCS  <NUM PROCESSES> valgrind --tool=massif --fair-sched=yes --massif-out-file=massif.%q{OMPI_COMM_WORLD_RANK} <executable> <executable args>

This will output massif.[rank] files. Run ms_print to convert these to a human readable format:

 ms_print massif.<rank> > massif.<rank>.log

Documentation

http://valgrind.org/docs/manual/manual.html