Algorithmic-Parameter Optimization of a Parallelized Split-Step Fourier Transform Using a Modified BSP Cost Model

摘要

Adaptive algorithms are increasingly acknowledged in leading parallel and distributed research. In the past, algorithms were manually tuned to be executed efficiently on a particular architecture. However, interest has shifted towards algorithms that can adapt themselves to the computational resources. A cost model representing the behavior of the system (i.e. system parameters) and the algorithm (i.e algorithm parameters) plays an important role in adaptive parallel algorithms. In this paper, we contribute a computational model based on Bulk Synchronous Parallel processing that predicts performance of a parallelized split-step Fourier transform. We extracted the system parameters of a cluster (upon which our algorithm was executed) and showed the use of an algorithmic parameter in the model that exhibits optimal behavior. Our model can thus be used for the purpose of self-adaption.