Article: KUMMS: optimising DRAM locality with Kernel-user behaviours Journal: International Journal of High Performance Systems Architecture (IJHPSA) 2016 Vol.6 No.1 pp.13 - 27 Abstract: DRAM accesses from operating system and user applications are analysed together in the existing researches, which try to improve the DRAM efficiency. In this paper, we find that Kernel and user perform different behaviours while accessing DRAM and interfere with each other greatly. Based on the observations, we divide the DRAM into Kernel-space, user-space and reserved-space. Different policies are designed for different spaces. We also introduce a new lock algorithm for the buddy system to deal with the page requests contentions in the multicore system. Experimental results show that the DRAM locality are averagely improved by more than 10% in the bank:row:column scheme and about 5% in the row:bank:column scheme. The improvements of RBH increase the system throughput without affecting the fairness, and averagely reduce the execution times of different benchmarks by 7.00% and 5.20% in the bank:row:column scheme and row:bank:column scheme, respectively. Inderscience Publishers - linking academia, business and industry through research

Title: KUMMS: optimising DRAM locality with Kernel-user behaviours

Authors: Beilei Sun; Xi Li; Chao Wang; Bo Wan; Xuehai Zhou

Addresses: School of Computer Science, University of Science and Technology of China, Hefei, Anhui, China ' School of Computer Science, University of Science and Technology of China, Hefei, Anhui, China ' School of Computer Science, University of Science and Technology of China, Hefei, Anhui, China ' School of Computer Science, University of Science and Technology of China, Hefei, Anhui, China ' School of Computer Science, University of Science and Technology of China, Hefei, Anhui, China

Abstract: DRAM accesses from operating system and user applications are analysed together in the existing researches, which try to improve the DRAM efficiency. In this paper, we find that Kernel and user perform different behaviours while accessing DRAM and interfere with each other greatly. Based on the observations, we divide the DRAM into Kernel-space, user-space and reserved-space. Different policies are designed for different spaces. We also introduce a new lock algorithm for the buddy system to deal with the page requests contentions in the multicore system. Experimental results show that the DRAM locality are averagely improved by more than 10% in the bank:row:column scheme and about 5% in the row:bank:column scheme. The improvements of RBH increase the system throughput without affecting the fairness, and averagely reduce the execution times of different benchmarks by 7.00% and 5.20% in the bank:row:column scheme and row:bank:column scheme, respectively.

Keywords: DRAM locality; kernel behaviour; user behaviour; row buffers; optimisation; page requests; multicore systems; operating systems; user applications.

DOI: 10.1504/IJHPSA.2016.076202

International Journal of High Performance Systems Architecture, 2016 Vol.6 No.1, pp.13 - 27

Received: 07 Aug 2015
Accepted: 09 Dec 2015
Published online: 29 Apr 2016 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: KUMMS: optimising DRAM locality with Kernel-user behaviours

Keep up-to-date