Theoretical analysis of classic algorithms on highly-threaded many-core GPUs

Abstract

The Threaded many-core memory (TMM) model provides a framework to analyze the performance of algorithms on GPUs. Here, we investigate the effectiveness of the TMM model by analyzing algorithms for 3 classic problems -- suffix tree/array for string matching, fast Fourier transform, and merge sort -- under this model. Our findings indicate that the TMM model can explain and predict previously unexplained trends and artifacts in experimental data.