基于InfiniBand的多链路mesh/torus大规模并行系统互连网络
InfiniBand-Based Multi-path Mesh/Torus Interconnection Network for Massively Parallel Systems
-
摘要: 在大规模并行系统中,系统级互连网络的设计至关重要.InfiniBand作为一种高性能交换式网络被广泛应用于大规模并行处理系统中.mesh/torus拓扑结构相较于目前普遍应用于InfiniBand网络的胖树拓扑结构拥有更好的性能与可扩展性.尽管如此,研究发现,用传统的mesh/torus拓扑结构构建InfiniBand互连网络存在诸多问题.分析了传统网络拓扑结构的缺陷,并提出了一种基于InfiniBand的多链路mesh/torus互连网络.这种改进型的拓扑结构通过充分利用交换机间的多链路可以获得比传统mesh/torus网络更高的带宽.另外,同时给出了与该网络拓扑结构相配套的高效路由算法.最后,通过网络仿真技术对提出的算法进行了评估,实验结果显示提出的路由算法相较于其他路由算法拥有更好的性能与可扩展性.Abstract: In the field of designing high performance computing system, the efficient implementation of interconnection network is mandatory because it has momentous influence on the performance of massively parallel computing systems. As a high performance switched interconnection network standard, InfiniBand is widely used in MPP systems. Compared with fat-tree topology which is commonly used in InfiniBand networks, mesh/torus topology can achieve better performance and scalability. However, according to our study, there are some challenges to use the traditional mesh/torus topology in building interconnection network. In this paper, we formulate the problem of traditional network topologies and propose an InfiniBand-based multi-link mesh/torus interconnection network which provides higher bandwidth than traditional networks by using multiple links between switches. In order to fully use the provided multi-link, the corresponding routing scheme for the improved network is also proposed. Simulation results show that the proposed routin scheme achieves better performance and scalability than other routing schemes.