Common MPI-Based Solutions for High-Performance Processing in Python Evaluated on Selected Test Cases [Presentation]

Presentation given in my thesis defense.

Abstract. A number of the most common high-performance computing approaches available in the Python programming environment of the LNCC Santos Dumont supercomputer are compared using three selected test cases. Python includes specific libraries, devel- opment tools, implementations, documentation and optimization or parallelization resources. It provides a straightforward way to allow programs to be written with a high level of abstraction, but the parallelization features to exploit multiple cores, processors or accelerators such as GPUs are diverse and may not be easily selectable by the programmer. This work compares common approaches in Python to increase computing performance. Three test cases are presented: a finite difference method for solving partial differential equations resulting from Poisson equations, a three- dimensional discrete Fourier transform method, and a random decision forest for ensemble learning method. The corresponding serial and parallel implementations in Fortran 90 were taken as references to compare their performance with some serial and parallel Python implementations of the corresponding algorithm. In addition to the performance results, a discussion of the trade-off between ease of programming and processing performance is included. This work is intended as a primer for using parallel HPC resources in Python.