面向共享Cache多核处理器的低功耗关键技术研究

项目名称： 面向共享Cache多核处理器的低功耗关键技术研究

项目编号： No.61202076

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： 方娟

作者单位： 北京工业大学

项目金额： 22万元

中文摘要： 低功耗是多核/众核处理器发展中所追求的重要目标之一。随着处理器核心数量的增多，会带来线延迟增加和功耗增大的问题。共享Cache的结构对多核体系结构的处理器功耗有着重要的影响。面向共享Cache的多核处理器低功耗技术的研究对未来不断发展的众核处理器、艾级超级计算机等的低功耗设计有重要意义。 本项目将在已有相关研究的基础上，着重对多核处理器中共享二级Cache的划分方法、路预测算法和Cache可重构方法进行研究来降低功耗。利用程序运行的局部性原理，通过私有和共享两种资源分配方式相结合来实施共享Cache的混合划分，通过关闭Cache列来降低功耗；通过综合考虑Cache的访问频率和LRU替换算法，预测同一个数据组中下一个预先访问的数据路，避免对不命中的路的访问；根据应用需求动态改变Cache的相联度及相应参数，使每一路或多路能在不同的情况下处于工作或休眠状态，从而降低功耗。

中文关键词： 多核处理器；低功耗；Cache划分；路预测；可重构

英文摘要： Low power is one important objective in developing multicore and manycore processor. With the increase of the number of cores, problems related to line-delay and energy consumption become increasing significant. The structure of a shared Cache is very important to the power management of multicore processors. Research on low-power multicore processors using shared Cache is of great significance to the design and development of manycore- and super-computer. The proposed project will conduct research on shared Cache partitioning based on way-prediction and reconfiguration to reduce power use of processors. In terms of the program locality principle, it employs both private and shared resource allocation methods to implement a hybrid partitioning strategy for power reduction. Taking into account access frequency and the LRU algorithm, it predicts the data way through which the same data set is accessed next and reduces the frequency of missed ways. Depending on the application context, corresponding associativity and parameters of the cache can be adjusted dynamically to reduce power use, by activating some ways and putting others to a dormant state.

英文关键词： Multicore processor；Low-power；Cache partitioning；Way-predicting；Cache reconfigurable