多核处理器目录缓存结构设计

发布时间：2018-07-24 20:40

【摘要】：随着物联网、云计算与网络舆情分析等应用的快速发展,大数据处理的应用已经成为数据中心的核心负载.数据中心服务器普遍采用多核处理器,而目录缓存作为多核处理器结构中维护缓存一致性的关键部件,对其结构研究(如稀疏目录)更多地关注于目录缓存的容量与可扩展性,更适合处理高性能计算等计算密集型应用.然而,当多核处理器执行延迟敏感的大数据应用程序时,目录缓存的高访存延迟严重制约了数据中心的服务质量.针对该问题,新型主从目录缓存结构优化了数据访问过程中的一致性协议通路,其中主目录区分共享与私有数据,管理私有数据的访存操作,降低私有数据的访存延迟,提高了从目录的容量利用率;从目录维护共享数据的缓存一致性,采用有限位标签结构,提高了从目录的存储效率.实验在Simics+GEMS模拟平台上对大数据程序测试集Cloudsuite-v1.0进行评估.结果表明在以大数据应用程序为主的运行环境下,与2倍容量的稀疏目录相比,主从目录缓存结构降低了24.39%的硬件开销,降低了28.45%的缓存缺失延时,提升了3.5%的处理器IPC;与缓存内目录相比,主从目录结构虽然损失了5.14%的缓存缺失延时与1.1%的处理器IPC,但是降低了42.59%的硬件开销.
[Abstract]:With the rapid development of Internet of things, cloud computing and network public opinion analysis, the application of big data processing has become the core load of data center. Multi-core processors are widely used in data center servers, and directory cache is the key component of maintaining cache consistency in multi-core processor architecture. Research on its structure (such as sparse directory) pays more attention to the capacity and extensibility of directory cache. More suitable for handling high performance computing and other computing intensive applications. However, when the multi-core processor executes a delay-sensitive big data application, the high memory access delay in the directory cache seriously restricts the quality of service of the data center. To solve this problem, the new master-slave directory cache structure optimizes the conformance protocol path in the process of data access, in which the home directory distinguishes the shared data from the private data, manages the private data access operation, and reduces the access delay of the private data. The capacity utilization of slave directory is improved, the cache consistency of shared data is maintained from directory to directory, and the storage efficiency of slave directory is improved by using finite bit tag structure. The Cloudsuite-v1.0 of big data program test set is evaluated on Simics GEMS simulation platform. The results show that the master-slave directory cache structure reduces the hardware overhead by 24.39% and the delay of cache deletion by 28.45% compared with the sparse directory with 2 times capacity in the big data application environment. Compared with the in-cache directory, the master-slave directory structure lost 5.14% cache deletion delay and 1.1% processor IPCs, but reduced the hardware overhead by 42.59%.
【作者单位】：高效能服务器和存储技术国家重点实验室(浪潮集团有限公司);
【基金】：国家“八六三”高技术研究发展计划基金项目(2013AA011701)
【分类号】：TP332

【共引文献】