Python and HDF5

Chapter 9. Concurrency: Parallel HDF5, Threading, and Multiprocessing

Over the past 10 years or so, parallel code has become crucial to scientific programming. Nearly every modern computer has at least two cores; dedicated workstations are readily available with 12 or more. Plunging hardware prices have made 100-core clusters feasible even for small research groups.

As a rapid development language, and one with easy access to C and FORTRAN libraries, Python is increasingly being used as a top-level “glue” language for such platforms. Scientific programs written in Python can leverage existing “heavy-lifting” libraries written in C or FORTRAN using any number of mechanisms, from ctypes to Cython to the built-in NumPy routines.

This chapter discusses the various mechanisms in Python for writing parallel code, and how they interact with HDF5.

Python Parallel Basics

Broadly speaking, there are three ways to do concurrent programming in Python: threads, the multiprocessing module, and finally by using bindings for the Message Passing Interface (MPI).

Thread-based code is fine for GUIs and applications that call into external libraries that don’t tie up the Python interpreter. As we’ll see in a moment, you can’t use more than one core’s worth of time when running a pure-Python program. There’s also no performance advantage on the HDF5 side to using threads, since the HDF5 library serializes all calls.

multiprocessing is a more recent built-in module available with Python, which provides support ...