A study of the single-program multiple-task model on GPU computing
Siqi Sun, Zhuo Zhang, Yanheng Zheng, Liang Wang, Weimin Xu, Wenfeng Shen · 2012
In recent years, GPU is rapidly developing in performance and used in wide range of extensive scientific and computation-intensive computing. However, the high performance on GPU depends on the degree of parallelism of applications. So GPU computing is usually used to solve a single large-size problem to utilize its massively parallel capability. In this paper, we devote the efforts to solve a number of variable-size tasks with the same algorithm on GPU and focus on the study of the Single-Program MultipleTask (SPMT) model on GPU computing. In addition, a novel allocation algorithm is proposed to map multiple tasks into GPU execution resources on SPMT model. The experiments show that our solution increase the performance of processing multiple tasks and can achieve up to 86.7% efficiency of single large-size task and 4.1 times speedup over sequential multi-task on Geforce 9800GTX+ in our example application.