MPI interface

The interface between the AMUSE python core and the legacy-codes is based on the MPI framework. Choosing MPI and not SWIG (or any other direct coupling method) has several advantages:

  • MPI is a well-known framework in the astrophysics community. Other coupling methods are less well known (like SWIG)

  • Legacy code does not run in the python space (memory usage, names)

  • Multiple instances of the same legacy code can easily be supported (not so in SWIG / f2py couplings)

  • Multi-process support taken into account at the start of the project.

  • Coupling is much looser.

There are also be some disadvantages:

  • Need to define a protocol over MPI

  • More “hand-work” needed to couple code. Other frameworks, like SWIG and f2py, generate an interface based on the application code.

  • More overhead for every call, slower calls

These disadvantages are mitigated by creating a library that handles most of the coupling details. This library has a Python, C++ and Fortran version. It implements the protocol and generates hooks to connect with the legacy codes.

The overhead per call may be an important factor in the speed of the framework. This will be tested during development of the first codes. It should be possible to limit the overhead by sending a lot of data per call. For example, setting the properties of a lot of stars in one call. Calling a lot of methods with limited data will be compared to sending one method with a lot of data.