Readme

Introduction

mpisee is a lightweight MPI profiler focusing on communicators. It can reveal communication bottlenecks and uncover hidden patterns within MPI communicators. Its output is an SQL database which can be queried for information using a python tool, mpisee-through.

Build

Dependencies

1. C/C++ compiler that supports C++11 standard
2. CMake version > 3.10
3. MPI Library (Open MPI, MPICH, CrayMPICH, MVAPICH, IntelMPI)
4. SQLite library

Basic build

git clone https://github.com/variemai/communicator_profiler.git && cd communicator_profiler
mkdir build && cd build
cmake -DCMAKE_INSTALL_PREFIX=/path/to/install .. && make -j && make install

• make install will place libmpisee.so in the lib folder of the installation path and mpisee-through.py to bin.
• You might need to define the compiler if it is not detected e.g., -DMPI_CXX_COMPILER=mpicxx.

Usage

• To profile your MPI application make sure it is compiled with the same MPI library as mpisee.
• Set LD_PRELOAD and MPISEE_OUTFILE variables before running your application.
• LD_PRELOAD to point to the libmpisee.so created in the build folder.
• MPISEE_OUTFILE defines output file name.
• The default output file name is “mpisee_” followed by date and time (e.g., mpisee_20240516143015.db).
• To get a summary of the results parse the output file with the mpisee-through.py.
  • For a complete list of options run /path/to/install/bin/mpisee-through.py --help.

Example Usage

LD_PRELOAD=/path/to/install/lib/libmpisee.so MPISEE_OUTFILE=/path/to/output.db <mpi launcher> <args> /path/to/exec <args>
/path/to/install/bin/mpisee-through.py -i /path/to/mpisee_profile.db

Results analysis with mpisee-through

• The default query displays all data in each communicator by summarizing across ranks:

  mpisee-through.py -i /path/to/mpisee_profile.db
      

• Display all data by separating ranks:

  mpisee-through.py -i /path/to/mpisee_profile.db -a
      

• Display data for collective MPI operations only:

  mpisee-through.py -i /path/to/mpisee_profile.db -c
      

• Display data for point-to-point MPI operations only:

  mpisee-through.py -i /path/to/mpisee_profile.db -p
      

• The following switches can be combined with the above options:
  • Display data for specific MPI ranks, e.g., Ranks 0 and 12: -r 0,12.
  • Display data for a specific buffer range, e.g., 0-1024: -b 0:1024.
  • Display data for MPI operations within a specific time range, e.g, from 0.02 to 0.8 seconds: -t 0.02:0.8.
• Use -h switch for a complete list of commands.

Check the wiki for a detailed example of how to interpret the output

Advanced Build

mpisee uses buffer size ranges to categorize the MPI communication calls. These buffer size categories can be configured by the user for finer grain analysis. The default values are: -NUM_BUCKETS=8 and -BUCKETS=128,1024,8192,65536,262144,1048576,33554432.

Buffer Range	Bucket
0-128	128
128-1024	1024
…	…

Configure Buffer Range Sizes (Buckets)

• Set the NUM_BUCKETS and BUCKETS correctly: length of BUCKETS must be equal to NUM_BUCKETS-1
• Sample build with 9 Buckets:

cmake -DNUM_BUCKETS=9 -DBUCKETS="64,512,4096,8192,65536,262144,1048576,33554432"

Cite

If you found this useful please cite: Link to IEEE Xplore

Bibtex

@INPROCEEDINGS{mpisee,
  author={Vardas, Ioannis and Hunold, Sascha and Ajanohoun, Jordy I. and Träff, Jesper Larsson},
  booktitle={2022 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW)},
  title={mpisee: MPI Profiling for Communication and Communicator Structure},
  year={2022},
  volume={},
  number={},
  pages={520-529},
  doi={10.1109/IPDPSW55747.2022.00092}}