面向大规模计算集群的多轨分割网络

发布时间：2018-04-20 05:21

本文选题：大规模计算集群 + 多轨网络　；参考：《计算机研究与发展》2017年11期

【摘要】：在千万亿次规模的系统中,互连网络设计面临新的挑战.高性能节点和大规模是构建千万亿次系统的主要技术趋势,不断提高的节点计算能力要求互连网络提供更高的性能,而不断增大的规模又对互连网络扩展性提出了更高的要求.此外,随着系统规模的增大,集合通信的执行时间也在不断增长,制约了应用的扩展性,集合通信的性能需要得到进一步优化.除性能之外,可靠性问题也随着系统规模的扩大而日益严重.而随着计算节点性能的不断提高,互连网络逐渐成为限制大规模计算机系统性能的瓶颈.互连网络核心部件交换芯片可提供的聚合网络带宽受到工艺和封装技术的限制.从网络结构与交换机结构的协同设计思想出发,提出了一种在交换机聚合带宽限定的条件下多轨分割网络结构和设计方法.通过数学建模和网络模拟仿真,分析了该多轨分割网络的性能边界.评测结果表明:该网络可将短消息(长度小于128B)的平均延迟性能提高10倍以上,为以短消息占多数的数据中心网络的性能优化提供了新思路.
[Abstract]:In the system of thousands of times scale, the design of interconnection network faces new challenges. High performance nodes and large scale are the main technical trends of constructing thousands of millions of times systems. The increasing node computing power demands higher performance of interconnection networks, and the increasing scale of interconnection networks puts forward higher requirements for the expansibility of interconnection networks. In addition, with the increase of system size, the execution time of collective communication is also increasing, which restricts the scalability of applications, and the performance of set communication needs to be further optimized. Besides performance, the problem of reliability becomes more and more serious with the expansion of system scale. With the continuous improvement of the performance of computing nodes, the interconnection network has gradually become the bottleneck that limits the performance of large-scale computer systems. The bandwidth of the converged network can be provided by the switching chip of the core component of the interconnection network, which is limited by the technology of process and encapsulation. Based on the cooperative design idea of network structure and switch structure, a multi-track partitioning network structure and design method under the condition of converging bandwidth of switch are proposed. The performance boundary of the multitrack segmented network is analyzed by mathematical modeling and network simulation. The evaluation results show that the network can improve the average delay performance of short message (length less than 128B) by more than 10 times, which provides a new way to optimize the performance of data center network with short message.
【作者单位】：计算机体系结构国家重点实验室(中国科学院计算技术研究所);中国科学院大学;
【基金】：国家重点研发计划项目(2016YFB0200300,2016YFGX030148,2016YFB0200205,2016GZKF0JT006) 国家自然科学基金项目(61572464,61331008,61402444) 国家“八六三”高技术研究发展计划基金项目(2015AA01A301) 华为科研基金项目(YB2015070066) 中国科学院战略性先导科技专项(XDB24060600)~~
【分类号】：TP393.0

【相似文献】