CPU-GPU融合架构上的缓存性能分析与优化

发布时间：2018-04-12 11:58

本文选题：异构架构 + 融合　；参考：《计算机工程与应用》2017年02期

【摘要】：现今CPU和GPU的发展已经出现新的瓶颈,将两者"结合"在同一块芯片上成为一种新的趋势。这种新的异构架构给片上共享资源的管理带来压力。而共享末级缓存(LLC)的管理对性能的影响非常关键。由于CPU程序和GPU程序的不同特性,给CPU和GPU间共享的末级缓存管理带来新的挑战。通过分析GPU程序访存特征,借鉴之前的缓存管理方案,提出对CPU-GPU融合系统的末级缓存进行等量的静态划分和最优静态划分的方案。实验结果表明:通过缓存划分可以有效避免CPU和GPU程序间的干扰。与传统LRU策略相比,等量静态划分和最优静态划分可以使系统整体性能分别提高7.68%和11.62%。
[Abstract]:Nowadays, the development of CPU and GPU has become a new bottleneck. It is a new trend to combine the two on the same chip.This new heterogeneous architecture puts pressure on the management of shared resources on a chip.The management of shared LLC is critical to performance.Because of the different features of CPU program and GPU program, it brings new challenges to the last level cache management shared between CPU and GPU.By analyzing the memory access characteristics of GPU programs and referring to the previous cache management schemes, this paper proposes a scheme of static partition and optimal static partition for the last level cache of CPU-GPU fusion system.Experimental results show that buffer partition can effectively avoid interference between CPU and GPU programs.Compared with the traditional LRU strategy, the static partition and optimal static partition can improve the overall performance of the system by 7.68% and 11.62%, respectively.
【作者单位】：中国科学技术大学计算机科学与技术学院;
【基金】：国家高技术研究发展计划(863)(No.2012AA010901,No.2012AA010902)
【分类号】：TP332

【相似文献】