面向多核NUCA共享数据竞争问题的Bank一致性技术  

Bank Coherence Techniques for Solving the Data Competition Problem in CMP NUCA

作　　者：吴俊杰[1] 潘晓辉[1] 

机构地区：[1]并行与分布处理国家重点实验室,湖南长沙410073

出　　处：《计算机工程与科学》2009年第11期21-24,49,共5页Computer Engineering & Science

基　　金：国家自然科学基金资助项目(60621003;60873014;60633050)

摘　　要：非一致Cache体系结构(NUCA)几乎已经成为未来片上大容量cache的发展方向。多核处理器的NUCA结构中,多个处理器核对共享数据的竞争访问,可能导致数据经常处于中部的cache Bank,增加NUCA的访问延迟。本文提出支持数据副本的Bank一致性技术,通过有选择地在NUCA中为访问的处理器核创建不同的数据副本,Bank一致性技术能够缓解多核处理器对共享数据的竞争问题。本文详细地介绍了Bank一致性协议的设计方法。最后,使用全系统模拟器对8个NPB基准测试程序进行了详细评测。实验结果表明,Bank一致性技术能够有效缓解多核处理器中共享数据的竞争访问问题。相比不支持Bank一致性技术的CMP-DNUCA结构,本文的方法能将系统IPC性能平均提升5.95%。Non-Uniform Cache Architecture （NUCA） has almost been the trend of large cache designs. In CMP-DNU- CA, two or more processors may access the same data, leading to the data competition problem. The data competition problem often makes the shared data stay in the central bank, so it brings a large cache access time. This paper proposes a bank coherence technique for supporting multi-copies of the shared data, which reduces the data competition effectively through making different data copies for processors. This paper studies the bank coherence protocol in detail. Finally, we test 8 NPB benchmarks using a full-system simulator. The experimental results show that the approach proposed in this paper effectively alleviates the data competition problem. Compared with CMP-DNUCA, the bank coherence mechanism achieves an average system IPC improvement of 5.95%.

关 键 词：非一致高速缓存 数据竞争 多核 存储体一致性 高速缓存一致性 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]