基于异构多核的LLC缓冲管理策略  

LLC Buffer Management Strategy Based on Heterogeneous Multi-core

作　　者：方娟 张希蓓 刘士建 FANG Juan;ZHANG Xibei;LIU Shijian(Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China;Beijing Intelligent City Insititute, Beijing University of Technology, Beijing 100124, China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]北京工业大学北京智慧城市研究院,北京100124

出　　处：《北京工业大学学报》2019年第5期421-427,共7页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(61202076);北京市自然科学基金资助项目(4192007)

摘　　要：针对异构多核架构下,大量共享末级缓存(last level cache,LLC)空间被GPU应用程序占据,导致系统性能降低的问题,提出了一种基于异构多核的LLC缓冲管理策略buffer-for-GPU(BFG).该策略在共享末级缓存旁侧建立一个与其结构相同的缓冲buffer,缓存不敏感GPU消息直接访问buffer而不访问LLC.针对CPU应用程序和GPU应用程序不同的特性,采用合理的替换策略,增加cache命中率.结果表明:BFG使CPU和GPU每一时钟周期内所执行的指令数(instruction per cycle,IPC)分别平均提升2. 48%和9. 30%,CPU应用程序命中率平均提升70. 45%. BFG可以有效地过滤GPU请求,为CPU应用程序腾出LLC空间,从而提高系统性能.Inheterogeneousmulti-corearchitecture,alargenumberofsharedlastlevelcache(LLC) space is occupied by GPU applications, leading to the system performance degradation. This paper presented a LLC buffer management strategy based on heterogeneous multi-core, buffer-for-GPU (BFG). A buffer was added on the side of LLC to filtrate streaming requests of GPU. Cache insensitive GPU messages directly accessed to buffer instead of accessing to LLC. Then, for different characteristics of CPU applications and GPU applications, a reasonable replacement strategy was adopted to increase the cache hit rate. Resulst show that compared with the traditional heterogeneous multi-core structure, the LLCbuffer management strategy improves the CPU’s and the GPU’s IPC by 2.48% and 9.30% on average,and improves the hitrate of CPU applications by 70.45% on average. BFG can effectively filter the GPU requests, vacate space to CPU applications, and improve system performance.

关 键 词：异构多核 缓存管理 未级缓存(LLC) 替换策略 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]