首页 | 本学科首页   官方微博 | 高级检索  
     

环连接CMP的缓存一致性协议
引用本文:曹非,刘志勇.环连接CMP的缓存一致性协议[J].计算机研究与发展,2009,46(Z2).
作者姓名:曹非  刘志勇
作者单位:1. 中国科学院计算技术研究所,北京,100190;中国科学院研究生院,北京,100049
2. 中国科学院计算技术研究所,北京,100190
基金项目:国家自然科学基金项目,国家"八六三"高技术研究发展计划基金项目,国家"九七三"重点基础研究发展计划基金项目 
摘    要:片上多核处理器(CMP)已经成为处理器发展的方向,处理器设计的重点也转到了互连网络和存储层次结构方面,其中的一个关键问题是如何维护各处理器各级缓存(Cache)的一致性,该问题在传统的共享存储多处理器中使用Cache一致性协议来解决,而CMP相对于传统的多处理器结构具有更高的片上互连带宽和速度,给Cache一致协议提出了新的要求,也提供了新的改进机会.传统的总线侦听协议存在可扩展性不足和不必要的广播、侦听过多的缺点,而目录协议则存在失效间接延时大和复杂度高、验证困难等问题.环形连接的可扩展性好于总线结构,而其实现复杂度也远小于通常目录协议所使用的包交换点到点网络.将基于环的侦听协议应用于CMP;并考虑利用环的顺序性取消原有协议中冲突引起的重发操作,消除可能的饥饿、死锁和活锁等情况,增加协议的稳定性,同时减少消息流量和功耗;利用片上互连延时短的特点,将侦听结果和侦听请求同时传播,使得处理器可以根据侦听结果来对侦听请求进行选择性的侦听操作,可减少不必要的侦听操作,降低功耗.

关 键 词:片上多核处理器  环连接  Cache一致性协议

Cache Coherence Protocol of Ring Connected CMP
Cao Fei,Liu Zhiyong.Cache Coherence Protocol of Ring Connected CMP[J].Journal of Computer Research and Development,2009,46(Z2).
Authors:Cao Fei  Liu Zhiyong
Abstract:Chip multiprocessor(CMP) is now a main direction of microprocessor research and development.On chip network and memory hierarchy are its design emphases,and one of the important problems is how to keep all the caches in coherence.In traditional shared memory multiprocessor architecture,cache coherence protocol is used to solve this problem. Compared with the traditional architecture,CMP has more on chip bandwidth and higher interconnection speed.This brings new challenges and improvement opportunities to the design of cache coherence protocol.Bus based snooping cache coherence protocol can't scale to systems with more processors and has many unnecessary broadcasts and snooping.Directory based protocol has long indirection latency,and is so complex to design and verify.Ring connection has more scalability than bus.while its complexity is far more less than the packet switch point-to-point interconnection network used by directory based systems.Snooping cache coherence protocol is applied on ring connected CMP,and ordering property of ring is used to cancel the retries of requests induced by colliding memory requests.Thus possible starvation,dead lock and live lock situations can be avoided,and the protocol has more performance stability and consumes less power and bandwidth.Because of the shorter latency of on chip links,the miss request and the snooping acknowledgement of it can be sent together,SO that following processors can use this information to decide whether to snoop or not.This can avoid unnecessary snooping and save power.
Keywords:chinp muttiprocessor(CMP)  ring  cache coherence protocol
本文献已被 万方数据 等数据库收录!
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号