一种面向多核处理器粗粒度的应用级Cache划分方法

A Coarse-Grand Cache Partitioning Technique for Multi-Programmed Workloads on CMPs

Cache划分技术是解决共享Cache访问冲突的重要方法,但是已有的Cache划分技术具有开销高、Cache划分时机难以确定的缺点。本文提出了面向应用的Cache划分框架(ACP)。ACP的优点是能够使用程序员提供的应用最外层循环的边界信息,更好地获取应用的失效率信息,因此Cache划分算法具有更高的精度,从而降低了划分的频率,进而提高系统性能。实验结果表明,和传统的固定周期的Cache划分方向相比,ACP具有更好的性能。

Cache partitioning is an effective way for reducing conflicted accesses from different processes to shared cache. However, existing cache partitioning techniques are difficult to determine the time partitioning shared cache. This paper presents application-level cache partitioning（ACP）. The partitioning time is determined by programmers based on the iteration boundary information, so it is easier for ACP to collect performance information and the cache partitioning algorithms are more effective with fewer cache partitioning and higher system performance. Experiment shows that ACP has higher performance.