一种监听均衡的片上多处理器Cache一致性协议  

Snooping-Balanced Cache Coherence Protocol for Chip Multiprocessors

作　　者：李永进[1] 邓让钰[1] 周宏伟[1] 

机构地区：[1]国防科技大学计算机学院,长沙410073

出　　处：《高性能计算技术》2012年第5期13-17,共5页

基　　金：本文获国家自然科学基金（61103011）和核高基重大专项（2009ZX01028-002-002）的资助.

摘　　要：基于目录的Cache一致性协议在片上多处理器（CMP：Chip Multiprocessor）中被广泛采用，监听是该协议中频繁使用的一类事务。过多地监听某个处理器核的私有Cache会干扰该核的正常读写操作，降低其访存带宽，这种现象随着CMP处理器中核数目的增加更加严重，造成各处理器核之间的访存带宽不平衡，影响同步操作的性能，降低处理器的并行性。在设计片上多处理器Cache一致性协议时，通过目录项中增加数据提供者标识，采用一种选择监听哪个处理器核的私有Cache以获取最新数据的策略，能够平衡各处理器核的私有Cache被监听打扰的概率，避免个别处理器核的正常访存操作被频繁干扰导致的性能显著下降，确保处理器的整体性能。Directory-based Cache coherence protocol are widely used today on chip multi-processor （CMP） and snoop is a class of transactions frequently used in CMP cache protocol. The normal read and write operations to one core will be disturbed if this core is accessed by snoop operations more frequently. It will cause that the memory bandwidth of this core is reduced. This problem become more serious with the increase of the number of cores in CMP and it resulting in that memory bandwidth between the processor cores is imbalance. The performance of synchronous operation is impacted and the parallelism of the processor is reduced. When CMP cache coherence protocol is designed, by adding the data provider information in directory entry and adopts a policy that which core＇s private cache is selected to be snooped in order to get the latest data, the frequency of each core＇s private cache is disturbed can be balanced. The performance decreased by the normal memory operations are interrupted frequently can be avoided, thus the whole performance of processor will be improved.

关 键 词：片上多处理器 CACHE一致性 监听 监听均衡 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]