Pakman

A modular, efficient and portable tool for running parallel approximate Bayesian computation algorithms.

Introduction

Pakman is a software tool for parallel approximate Bayesian computation (ABC) algorithms. Its modular framework is based on user executables, which means that problem-specific tasks, like model simulations, are performed by black box executables supplied to Pakman by the user.

Pakman parallelises the execution of simulations using MPI, a portable standard for distributed computing, and was designed to be lightweight so that a minimal amount of overhead goes into parallelisation.

The problems that will benefit the most from Pakman are those where model simulations take a relatively long time, on the order of seconds or more.

Requirements

• C++11 compiler
• MPI-3.1 library
• CMake >= version 3.10.2
• (Optional) Python >= version 3.6.7 with numpy, matplotlib, and scipy

Pakman has been tested with OpenMPI and MPICH. Python is not necessary to build Pakman, but it is used in some Pakman examples and to create figures.

Building

To build Pakman, run:

$ mkdir build $ cd build $ cmake .. $ make

This will create a pakman binary under the src subfolder of the build directory.

Testing

To test Pakman, run (in the build folder):

$ ctest

This will run a series of tests to verify that the Pakman build is working correctly.

To test how well Pakman scales with the number of parallel processes employed, run (in the build folder):

$ scaling/run-scaling.sh

This script will benchmark Pakman with a computationally intensive simulator for different numbers of parallel instances of the simulator. The results are saved in the comma-separated file scaling.csv. In addition, if Python was detected by CMake, the speedup and efficiency with respect to the number of processes will be plotted in speedup.png and efficiency.png, respectively.

It is recommended to use the flag -DCMAKE_BUILD_TYPE=Release with the cmake command before running the scaling test to reduce computation time.

Documentation

Examples of how to use Pakman can be found in the folder examples inside the build folder. See the wiki for detailed documentation.

• Biased coin and ABC rejection
• Biased coin and ABC SMC
• SIS model
• Epithelial cell growth

In addition, the command build/src/pakman --help provides a quick reference on how to use Pakman.

Developers: code documentation can be found here.

Testing on multiple nodes

By default, running ctest will only test Pakman on the local node. Since Pakman uses MPI for parallelisation however, it is possible to run Pakman on multiple nodes. In order to run the above tests and examples on the nodes that specific to your setup, you need to pass additional information to CMake.

Firstly, you need to define the CMake variable MPIEXEC_HOSTS_FLAGS to contain the command-line flags you would pass to mpiexec to specify the hosts. Secondly, define MPIEXEC_MAX_NUMPROCS to specify the total number of MPI processes to run.

For example, if you would normally launch 8 MPI processes on node0 and node1 in the following manner:

$ mpiexec --hosts node0,node1 -n 8 ...

Then you need to run CMake as:

... $ cmake .. -DMPIEXEC_HOSTS_FLAGS="--hosts node0,node1" -DMPIEXEC_MAX_NUMPROCS=8 ...

When building Pakman with these flags, CMake will automatically insert the appropriate flags in the mpiexec commands for running tests (including the scaling benchmark) and examples. Thus, running the above commands as before will now test Pakman on multiple nodes.

Contributing

We welcome your contributions! Please see our contributing guidelines.