mpiprof 0.2.2

Installation

• From pypi:

  pip install mpiprof

• From a local checkout, by cloning this repository:

  git clone https://github.com/JohannesBuchner/mpiprof.py
  cd mpiprof.py/
  pip install .

Requirements

• Python 3.8+

• mpi4py

• An MPI runtime if you plan to launch under mpiexec/srun/etc.

Usage: per-rank cProfile runner

Run your script under mpiexec (or srun) using the module form:

• mpiexec -n 4 python3 -m mpiprof your_script.py arg1 arg2

This writes one profile file per rank named mpiprof.<rank>.pstats in the current working directory.

Options:

• -o / --outfile sets the output path or pattern (default: mpiprof.{rank}.pstats). The literal substring {rank} will be replaced with the MPI rank.

Signal handling:

• mpiprof installs SIGINT and SIGTERM handlers to dump the profile before exiting, so Ctrl-C or a clean termination should still produce output. If the MPI launcher escalates to SIGKILL immediately, no tool can save a profile.

Usage: MPIProfiler for mpi4py

The MPIProfiler wraps a communicator and measures wall-clock time and call counts for common blocking MPI calls. It is intentionally lightweight and safe to leave in production runs (low overhead on the measured calls). Nonblocking calls are counted but not timed precisely (the time recorded is call overhead, not the transfer time).

Example:

from mpi4py import MPI
from mpiprof import MPIProfiler

comm = MPIProfiler(MPI.COMM_WORLD)

# Your MPI code as usual, using comm instead of MPI.COMM_WORLD:
rank = comm.Get_rank()
size = comm.Get_size()
data = rank

# Simple collective
total = comm.allreduce(data, op=MPI.SUM)

# Point-to-point
if rank == 0:
    comm.send(b"hello", dest=1, tag=0)
elif rank == 1:
    msg = comm.recv(source=0, tag=0)

# Write stats at the end (one file per rank)
comm.write_statistics()  # default name: mpiprof.stats.<rank>.json

Notes:

• The wrapper exposes common methods (e.g., send, recv, bcast, reduce, allreduce, gather, scatter, barrier) and forwards any other attributes to the underlying communicator. It tries to be case-insensitive to match mpi4py idioms (both Send and send are supported).

• For nonblocking operations (Isend, Irecv), the wrapper records the call count but cannot attribute data transfer time unless you also wrap and time Wait/Waitall. A simple wait wrapper is provided to time individual requests returned by the wrapper.

Rank detection

The runner tries to detect the rank via common environment variables:

• OMPI_COMM_WORLD_RANK, PMIX_RANK (Open MPI)

• PMI_RANK (MPICH, Intel MPI)

• MV2_COMM_WORLD_RANK (MVAPICH2)

• SLURM_PROCID (scheduler fallback)

• default: 0 if none found

Output files

• Runner: mpiprof.<rank>.pstats (or the pattern you set with -o). You can analyze it with pstats or tools like snakeviz:

  • python3 -m pstats mpiprof.0.pstats

  • snakeviz mpiprof.0.pstats

• MPIProfiler: MPIprofile.<rank>.out with operation counts and total wall-clock time per operation. Example:

  Function: scatter
  Call stack:
          surveymcmc.py:466
          surveymcmc.py:431 sampler.advance_one_step(False)
          mcaeis.py:107 local_coords = self.comm.scatter(chunks, root=0)
  Number of calls: 2
  Duration During Call: 98.990435s
  Duration Before Call: 0.000014s
  
  Function: scatter
  Call stack:
          surveymcmc.py:466
          surveymcmc.py:448 sampler.advance_one_step()
          mcaeis.py:75 self._init()
          mcaeis.py:55 self.log_probs = self._evaluate_log_probs(self.coords)
          mcaeis.py:64 local_coords = self.comm.scatter(chunks, root=0)
  Number of calls: 39
  Duration During Call: 2.228181s
  Duration Before Call: 0.000025s
  
  Total MPI Time: 388.922329s
  Total Non-MPI Time: 206.503627s

Results are sorted (descending) by duration during call.

Limitations

• The runner cannot save profiles if the process is killed by SIGKILL.

• MPIProfiler’s accounting for nonblocking calls is approximate unless you consistently call wait/waitall on the requests returned by the wrapper’s nonblocking methods.

License

MIT