An Efficient Data Prefetching Scheme in GPUs

Please enable javascript in your browser.

Mostofi, Saba | 2023

83 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 56200 (19)
University: Sharif University of Technology
Department: Computer Engineering
Advisor(s): Sarbazi Azad, Hamid
Abstract:
GPUs exploit memory hierarchy and thread-level parallelism (TLP) to hide off-chip memory access delay. However, GPUs can not keep TLP high during the execution of various applications, and hence, they fall short of hiding the access delay to off-chip memory. One effective approach to reducing memory access delay is prefetching. Prior research shows the positive impact of prefetching in improving the performance of GPUs. However, they fail to capture all the potential behind the prefetching in GPUs. In this thesis, we propose Snake, a new data prefetching in GPUs. Snake identifies stride distances among different memory access instructions and prefetches a chain of addresses that will be needed by threads in the future. We explore the challenges and devise a mechanism to perform the chain-based prefetching effectively. The proposed method can prefetch 80% of memory access requests with an accuracy of 90%. Evaluation results show that, Snake improves the performance of GPUs by 16% and reduces energy consumption by 17% over 11 GPU applications
Keywords:
Graphic Processing ; Data Prefetching ; On-Chip Memories ; Performance Evaluation ; Hardware Prefetching