Abstract: Based on the statistical analysis of eight benchmarks simulation, this paper proposes selective replication for some read-only shared data, and in-situ communication for write-read shared data. The former means to improve cache average access speed, while the latter’s purpose is to reduce coherence overhead. It extends a more simple method to solve the ping-pong problem. Statistical analysis of simulation shows these two schemes achieve less access latency and better space utilization compared with conventional cache scheme.

Key words: Symmetric Multiprocessor(SMP), selective replication, cache

摘要： 在统计分析8个典型测试程序的模拟运行的基础上，提出在多核处理器的私有L1上对部分只读共享数据进行复制以加快访问速度，对读-写共享数据采用“原地通信”策略，以减少一致性开销，针对其中的“写无效化共享标记”问题提出一种更简单的解决方法。从模拟实验的统计分析可知，采用这2种策略可获得比传统策略更小的平均访问时延和更高的空间利用率。

关键词: 对称多处理器, 选择性复制, 高速缓冲

CLC Number: 

• TP333