A compiler-assisted runtime-prefetching scheme for heterogeneous platforms

  • Authors:

  • Li Chen;Baojiang Shou;Xionghui Hou;Lei Huang

  • Affiliations:

  • State Key Laboratory of Computer Architecture, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China;State Key Laboratory of Computer Architecture, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China;State Key Laboratory of Computer Architecture, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China;Department of Computer Science, Prairie View A&M University, Prairie View, TX

  • Venue:
  • IWOMP'12 Proceedings of the 8th international conference on OpenMP in a Heterogeneous World
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

GPGPU has been widely used in recent years in both academia and industry. Many research for benchmarks on GPUs were reported to achieve over 100 times speedup, however, due to the high overhead of data transfer between GPU and CPU in real-world applications, the achievements are dramatically limited. In the case of using multiple GPUs, the situation is even worse. Another difficulty raised by the GPGPUs is the programming productivity. In this work, we introduce a new language extension to the easy-to-use programming model OpenMP, implement a runtime and a prefetching mechanism to further extend our work in support of OpenMP on heterogeneous platforms. The new language extension allows the OpenUH compiler to generate efficient code for heterogeneous platforms with multiple GPUs included. To improve the performance of applications with lots of data transfer, we implement runtime inter-thread dataflow analysis and a runtime-prefetching mechanism with the help of compiler analysis, making the data transfer overlap with the computation as much as possible. We have evaluated our prefetching system using benchmarks including NPB SP, kmeans and nbody. In these experiments, we achieve speedups of 1.23, 1.4 and 1.32 respectively compared with the versions without prefetching support.