高性能计算机片上网络交换结构和性能的研究

发布时间：2018-04-05 23:00

本文选题：高性能计算机　切入点：片上网络　出处：《北京邮电大学》2012年硕士论文

【摘要】：在目前的高性能计算系统中,为了弥补单个处理器计算能力的不足,需要将多个处理器互连,进行并行计算,从而提高整个系统的计算能力。在目前的高性能计算系统电互连片上网络中,电互连的一些固有缺陷成为了抑制网络性能提高的主要障碍。与电互连相比,片上光互连具有更好的性能,是芯片级互连中更好的解决方案。为了突破基于电互连的片上网络的瓶颈,满足日益增加的对高性能计算技术的需求,本文对片上光网络进行了研究。本论文完成的主要工作如下： 1,完成了一种基于簇状网络拓扑结构(Cluster Mesh)的片上光网络的研究。对该网络的网络结构、节点结构、通信协议和路由策略进行了设计,并使用OPNET网络仿真工具进行了仿真建模。通过仿真,研究了该簇状网络的网络性能。将仿真结果与传统的Mesh网络仿真结果进行比较,并分析了簇状网络在局部流量较大时所具有的优越性。 2,对一种无缓存的基于分组交换的光网络(BOIN)进行了仿真研究。对该网络的网络结构和路由方式进行了理论研究,使用OPNET软件进行了仿真建模,并通过仿真分析了网络的时延和吞吐量等相关性能参数。通过将BOIN网络与传统的基于电路交换的Torus网络的仿真性能进行对比,体现了BOIN网络在网络资源利用上的优势。通过研究和仿真,我们得到了簇状网络拓扑结构和分组交换这两项技术在芯片级的光网络上的性能优势。仿真结果表明,在网络局域流量较大时,簇状网络拓扑结构比相同规模的普通Mesh拓扑结构具有更好的时延和吞吐量性能。而基于分组交换的片上光网络也比基于电路交换的片上网络具有更高的链路利用率。这两项技术也将成为芯片级光互连网络未来的研究热点。
[Abstract]:In the current high performance computing system, in order to make up for the lack of computing power of a single processor, it is necessary to interconnect multiple processors for parallel computing, so as to improve the computing power of the whole system.In the current high performance computing system, some inherent defects of electrical interconnection have become the main obstacle to restrain the improvement of network performance.Compared with electrical interconnection, on-chip optical interconnection has better performance and is a better solution in chip level interconnection.In order to break through the bottleneck of on-chip network based on electrical interconnection and meet the increasing demand for high performance computing technology, the optical on-chip network is studied in this paper.The main work of this thesis is as follows:1. A novel on-chip optical network based on cluster Mesh topology is proposed.The network structure, node structure, communication protocol and routing strategy of the network are designed, and the simulation model is built by using OPNET network simulation tool.The network performance of the cluster network is studied by simulation.The simulation results are compared with the traditional Mesh network simulation results, and the advantages of the cluster network when the local traffic is large are analyzed.2. A buffer-free optical network based on packet switching is simulated.The network structure and routing mode of the network are studied theoretically, and the simulation model is built by using OPNET software, and the network performance parameters such as delay and throughput are analyzed by simulation.By comparing the simulation performance of BOIN network with that of traditional Torus network based on circuit switching, the advantage of BOIN network in network resource utilization is demonstrated.Through research and simulation, we obtain the performance advantages of cluster network topology and packet switching in chip optical networks.The simulation results show that the cluster topology has better delay and throughput performance than the common Mesh topology with the same size when the local traffic is large.On-chip optical networks based on packet switching have higher link utilization than on-chip networks based on circuit switching.These two technologies will also become the future research hotspot of chip-level optical interconnection network.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TN47;TP338

【相似文献】