Memory Request Priority Based Warp Scheduling for GPUs  

作　　者：ZHANG Jun HE Yanxiang SHEN Fanfan LI Qing'an TAN Hai 

机构地区：[1]School of Software,East China University of Technology [2]Computer School,Wuhan University

出　　处：《Chinese Journal of Electronics》2018年第5期985-994,共10页电子学报（英文版）

基　　金：supported by the National Natural Science Foundation of China(No.61662002,No.61373039,No.61462004);the Science and Technology Project of the Educational Deparment in Jiangxi Province,China(No.GJJ150605);the Natural Science Foundation of Jiangxi Provice(No.20151BAB207042,No.20161BAB212056);the Key Research and Developement Plan of the Scientific Department in Jiangxi Province,China(No.20161BBE50063)

摘　　要：High performance of GPGPU comes from its super massive multithreading, which makes it more and more widely used especially in the field of throughputoriented. Data locality is one of the important factors affecting the performance of GPGPU. Although GPGPU can exploit intra/inter-warp locality by itself in part, there is still large improvement space for that. In our work, we analyze the characteristics of different applications and propose memory request based warp scheduling to better exploit inter-warp spatial locality. This method can make some warps with good inter-warp locality run faster, which is beneficial to improve the whole performance. Our experimental results show that our proposed method can achieve24.7% and 11.9% average performance improvement over LRR and MRPB respectively.High performance of GPGPU comes from its super massive multithreading, which makes it more and more widely used especially in the field of throughputoriented. Data locality is one of the important factors affecting the performance of GPGPU. Although GPGPU can exploit intra/inter-warp locality by itself in part, there is still large improvement space for that. In our work, we analyze the characteristics of different applications and propose memory request based warp scheduling to better exploit inter-warp spatial locality. This method can make some warps with good inter-warp locality run faster, which is beneficial to improve the whole performance. Our experimental results show that our proposed method can achieve24.7% and 11.9% average performance improvement over LRR and MRPB respectively.

关 键 词：General purpose graphics processing unit（GPGPU） Memory request priority Performance Data locality Warp scheduling 

分 类 号：TP333[自动化与计算机技术—计算机系统结构]