片上多处理器共享Cache优化关键技术研究

项目名称： 片上多处理器共享Cache优化关键技术研究

项目编号： No.61202126

项目类型： 青年科学基金项目

立项/批准年度： 2013

项目学科： 计算机科学学科

项目作者： 徐金波

作者单位： 中国人民解放军国防科学技术大学

项目金额： 24万元

中文摘要： 片上多处理器通常采用共享最后一级Cache的方式。由于多核对共享Cache的竞争问题，同时兼顾多核对Cache访问的高性能和公平性具有较大挑战性。针对此，从多个角度进行Cache优化，并有机整合为完整的Cache优化框架。 在硬件结构方面，借助理论分析与实验验证手段研究共享Cache大小、数量、组织方式、互连结构对线延迟的影响，通过物理路径优化和逻辑距离优化，缩短线延迟；在资源管理方面，研究基于各个核的负载分布情况进行公平性度量的共享Cache资源动态公平划分机制；在数据调度方面，研究充分挖掘核间数据访问模式协同性的数据预取策略；在失效替换策略方面，研究结合数据访问频率信息、实现机制更简单的替换策略。 通过有机集成以上优化策略，能够降低硬件结构复杂性，减小访问延迟，动态保证资源划分公平性，增强Cache与处理器核的协同性，提高命中率和资源利用率，实现多核性能的均衡提升。

中文关键词： 片上多处理器；共享缓存；资源划分；数据预取；替换策略

英文摘要： Chip Multiprocessors usually share last-level on-chip Caches. Due to contention of cache resources shared among multiple cores, it is challenging to achieve both high performance and fairness for per-core's accessing shared cache. Aiming at this problem, this project focuses on cache optimization strategies from multiple aspects. These strategies will be integrated into an entire optimization framework organically. From the hardware infrastructure aspect, this project researches how the size, bank numbers, organization fashion and interconnect structures of shared cache impact wire delays by using theoretical analysis and experimental verification. The wire delays are expected to be shortened both physically and logically. From the resource management aspect, dynamic fair cache sharing and partitioning scheme based on workload distribution patterns of different cores are studied. From the data scheduling aspect, this project focuses on data prefetching strategy which fully exploits commonalities of data access patterns among different cores. From replacement policy aspect, the data access frequency information will be incorporated into the decision-making of replacement, and the implementation cost and complexity are expected to be small. By integrating the above strategies, the resulting cache optimization fram

英文关键词： Chip Multiprocessor；Shared Cache；Resource Partitioning；Data Prefetching；Replacement Policy