MPD:结点具有多个并行缓存一致性域的CC-NUMA系统
发布时间：2019-05-26 18:49
【摘要】：大规模高速缓存一致性非均匀存储访问(cache coherence non-uniform memory access,CC-NUMA)系统通常采用两级一致性域方法来降低缓存一致性协议维护开销,提升系统性能.两级一致性域系统中,多个处理器互连,形成结点内一致性域;多个结点互连,形成结点间一致性域.然而,受限于处理器直连能力与处理器可识别ID数,系统的单结点规模有限,系统规模的扩展不得不依靠增加结点数来实现,使得大规模CC-NUMA系统的结点间互连复杂度上升,跨结点访问带宽和延迟急剧增长,影响了系统性能的有效扩展.MPD系统通过在结点内构建多个并行缓存一致性域,突破了处理器直连能力与可识别ID数对单结点规模的限制,能够大幅减少结点数量,并将部分结点间访问转化为结点内访问,实现系统性能的有效扩展.理论分析和实验结果表明:采用同规格处理器的32路系统中,结点内4个并行缓存一致性域的MPD系统可实现结点数目减少75%、一致性目录存储开销节省40%以上、平均访问延迟降低约27.9%、系统整体性能提升约14.4%.
[Abstract]:Large-scale cache consistent non-uniform storage access (cache coherence non-uniform memory access,CC-NUMA) systems usually use a two-level consistency domain approach to reduce cache consistency protocol maintenance overhead and improve system performance. In a two-level uniform domain system, multiple processors interconnect to form a consistent domain within a node, and multiple nodes interconnect to form a consistent domain between nodes. However, limited by the direct connection ability of the processor and the number of ID identifiable by the processor, the single node size of the system is limited, and the expansion of the system scale has to be realized by increasing the number of nodes, which increases the complexity of interconnection between nodes in large-scale CC-NUMA systems. The rapid increase of cross-node access bandwidth and delay affects the effective expansion of system performance. MPD system breaks through the limitation of processor direct connection ability and identifiable ID number to single node size by building multiple parallel cache consistency domains in nodes. The number of nodes can be greatly reduced, and the access between some nodes can be transformed into intra-node access, so as to realize the effective expansion of system performance. The theoretical analysis and experimental results show that in the 32-channel system with the same specification processor, the number of nodes can be reduced by 75% and the storage cost of the consistent directory can be saved by more than 40% in the MPD system with four parallel cache consistent domains in the node. The average access delay is reduced by about 27.9%, and the overall performance of the system is improved by about 14.4%.
【作者单位】： 高效能服务器和存储技术国家重点实验室(浪潮集团有限公司);
【分类号】：TP303
本文编号：2485545
[Abstract]:Large-scale cache consistent non-uniform storage access (cache coherence non-uniform memory access,CC-NUMA) systems usually use a two-level consistency domain approach to reduce cache consistency protocol maintenance overhead and improve system performance. In a two-level uniform domain system, multiple processors interconnect to form a consistent domain within a node, and multiple nodes interconnect to form a consistent domain between nodes. However, limited by the direct connection ability of the processor and the number of ID identifiable by the processor, the single node size of the system is limited, and the expansion of the system scale has to be realized by increasing the number of nodes, which increases the complexity of interconnection between nodes in large-scale CC-NUMA systems. The rapid increase of cross-node access bandwidth and delay affects the effective expansion of system performance. MPD system breaks through the limitation of processor direct connection ability and identifiable ID number to single node size by building multiple parallel cache consistency domains in nodes. The number of nodes can be greatly reduced, and the access between some nodes can be transformed into intra-node access, so as to realize the effective expansion of system performance. The theoretical analysis and experimental results show that in the 32-channel system with the same specification processor, the number of nodes can be reduced by 75% and the storage cost of the consistent directory can be saved by more than 40% in the MPD system with four parallel cache consistent domains in the node. The average access delay is reduced by about 27.9%, and the overall performance of the system is improved by about 14.4%.
【作者单位】： 高效能服务器和存储技术国家重点实验室(浪潮集团有限公司);
【分类号】：TP303
【相似文献】
相关会议论文 前1条
1 董小洁;夏宽理;;忍受犯规的开放式提交协议[A];第十二届全国数据库学术会议论文集[C];1994年
相关重要报纸文章 前1条
1 蒋杰 方力 窦文华;覆盖控制[N];计算机世界;2004年
相关博士学位论文 前1条
1 王立;城市结点文化特质及其协同观[D];重庆大学;2006年
相关硕士学位论文 前6条
1 杨孟君;基于网络认知的无中心式系统交互的优化方法[D];电子科技大学;2015年
2 田艳;复杂网络重要结点发现算法研究[D];云南财经大学;2016年
3 姜浩;对等网络中路由中继结点发现机制的研究[D];华中科技大学;2007年
4 顾烨;P2P网络逻辑拓扑优化和结点组管理策略研究[D];浙江工商大学;2010年
5 李丁丁;一种基于结点聚类的网络定位算法[D];湖南大学;2008年
6 孙宇奇;基于复杂网络的社团发现研究[D];辽宁师范大学;2011年
，本文编号：2485545
本文链接：https://www.wllwen.com/kejilunwen/jisuanjikexuelunwen/2485545.html
