基于CMP的共享L2Cache管理策略研究

发布时间：2018-06-09 14:51

本文选题：片上多核处理器 + 共享L2Cache划分　；参考：《哈尔滨工程大学》2013年硕士论文

【摘要】：随着多核处理器体系结构的飞速发展，片上多核处理器以其特有的优势成为主流多核处理器架构被广泛应用到服务器、PC机等商品中，对其进行的研究也在不断深入。多核处理器在多任务处理、高性能计算方面优势显著，但还存在着制约其性能发挥的因素，如存储墙问题在多核处理器系统中依然严重，极大地影响了多核处理器系统总体性能的进一步提升。针对以上问题，本文将从共享L2Cache资源划分技术和应用程序的L2Cache访问类型感知的数据块替换策略两个方面具体阐述基于CMP的共享L2Cache管理策略。首先，，在片上多核处理器的基础上，针对现有划分技术中普遍忽略的失效开销差异性问题，提出了一种基于访存时间的L2Cache公平性划分策略，该策略通过L2MSHR、监测器和HSHR三种硬件配置获取失效率并行度信息计算各线程在独占和共享L2Cache两种情况下运行时间，并在MTFP算法和公平性评价指标FairT的指导下获得每个划分周期内各线程的分配路数。同时，针对LRU策略在大工作集应用程序和时间局部性差的流式应用程序中存在的性能严重下降的问题，提出了一种基于L2Cache访存访问类型感知的替换策略，该策略在ADC算法的指导下感知各个应用程序在不同时间周期内体现出的不同L2Cache访存类型动态调整提升和插入方式，有效地降低了低频访问数据块和零重用数据块在L2Cache的驻留时间。最后，结合MTFP策略和RPBL策略提出L2Cache管理策略LMSC，并对LMSC策略的运行机制进行了具体阐述。通过基于Gems和Simics的性能评价方案，从吞吐率、公平性和加权加速比三个方面对LMSC策略进行测评。实验结果表明：与LRU策略相比，LMSC策略能够在保证系统吞吐率的同时提升了加权加速比和系统公平性，具有良好的研究价值。
[Abstract]:With the rapid development of multi-core processor architecture, multi-core processor on chip has become the mainstream multi-core processor architecture because of its unique advantages. Multi-core processors have significant advantages in multi-task processing and high-performance computing, but there are still some factors that restrict their performance. For example, the problem of storage wall is still serious in multi-core processor systems. Has greatly affected the further improvement of the overall performance of multi-core processor systems. This paper describes the shared L2Cache management strategy based on CMP from two aspects: shared L2Cache resource partition technology and L2Cache access type aware data block replacement strategy of application program. Firstly, on the basis of multi-core processor on chip, a L2Cache fairness partitioning strategy based on memory access time is proposed, which is generally ignored in the existing partition technology. Through L2MSHRR, monitor and HSHR hardware configuration, the strategy acquires the failure rate parallelism information to calculate the running time of each thread in the case of exclusive and shared L2Cache. Under the guidance of MTFP algorithm and FairT, the number of allocation paths for each thread in each partition cycle is obtained. At the same time, a replacement strategy based on L2Cache access type awareness is proposed to solve the problem of serious degradation of LRU policy performance in large working set applications and streaming applications with poor temporal locality. Under the guidance of ADC algorithm, this strategy is aware of the different L2Cache memory access types dynamically adjusted and inserted in different applications in different time cycles. The low frequency access data block and the zero reuse data block residence time in L2Cache are reduced effectively. Finally, L2Cache management policy LMSCS is proposed by combining MTFP policy with RPBL policy, and the running mechanism of LMSC policy is expounded in detail. Through the performance evaluation scheme based on Gems and Simics, the LMSC strategy is evaluated from three aspects: throughput, fairness and weighted speedup. The experimental results show that compared with LRU strategy, LMSC strategy can improve the weighted speedup and system fairness while ensuring the throughput of the system, and has good research value.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP332

【参考文献】