Parallelism vs. speculation

Abstract

Graphics Processing Unit (GPU) shows stunning computing power for scientific applications in the past few years, which attracts attention from both industry and academics. The huge number of cores means high parallelism and also powerful computation capacity. Many previous studies have taken advantage of GPU's computing power for accelerating scientific applications. The common theme of those research studies is to exploit the performance improvement provided by massive parallelism on GPU. Despite that there have been fruitful research work for speeding up scientific applications, little attention has been paid to the redundant computation resources on GPU. Recently, the number of cores integrated in a single GPU chip increases rapidly. For example, the newest NVIDIA GTX 980 device has up to 2048 CUDA cores. Some scientific applications, such as Genetic Algorithm (GA), may have an alternative way to further improve their performance. In this paper, based on the biological fundamentals of GA, we propose a speculative approach to use the redundant computation resources (i.e., cores) to improve the performance of parallel genetic algorithm (PGA) applications on GPU. Comparing to the traditional parallelism scheme, our theoretical analysis shows that the speculative approach should improve the performance of GA applications intuitively. We experimentally compare our design with the traditional parallelism scheme on GPU using three Nonlinear Programming problems (NLP). Experimental results demonstrate the effectiveness of our speculative approach in both execution time and solution accuracy of GA applications on GPU.