Design And Evaluation Of High-radix Topology In High Performance Interconnection Network

No evidence has indicated the end of the greedy demand of higher performance,faster speed transmission and larger scale of high performance computing systems in the information era of today. The rapid development of microelectronics technology has greatly boosted the bandwidth and enhanced the links of a single router chip. The traditional topology for interconnection networks cannot conform with the increasing interconnection bandwidth and density requirement of on-chip and off-chip interconnection any more due to its inherent physical properties, and make itself a bottleneck to the performance of high performance computing systems. Facing the emergence of the interconnection crisis and the opportunity of router chip, engineers have to consider high performance computing systems from the point of interconnection design. How to utilize high-radix router to reduce the latency and cost is the key for high performance network topology. Based on the above problems, this paper studies the content from the following four aspects:(1)Based on an in-depth understanding of the high-radix interconnection network and the current development of high-radix router, we summarize the impact of high-radix router to the design of topology for high performance interconnection network and the limitations of low-radix interconnection network to using high-radix routers.(2) As for high-radix topology for high performance interconnection network status quo, we focus on three typical high-radix topology for high performance interconnection network: Fat tree, Flattened Butterfly and Dragonfly. These three topologies utilize the properties of high-radix routers and traditional low-radix topologies, such as k-ary n-cube,Butterfly and so on. Apart from the theoretical analysis, the interconnection network simulator xNetSim is adopted for evaluation. These three topologies can not only support for larger networks but also effectively reduce the diameter and network latency. However,these three topologies remain inadequate of scalability, bisections, overhead and other indicators.(3)To improve the typical high-radix topology, we propose a new scalable topology SuperStar for high performance interconnection network. The SuperStar, whose diameter and cost are effective. Meanwhile, SuperStar takes advantages of high-radix routers to approve of larger systems. Apart from theoretical analysis, we explore the high-radix interconnection network simulator xNetSim to evaluate it. In the simulations, we vary the network loads and verify the throughput and latency of different routing algorithms.In addition, we briefly analyze the performance of SuperStar comparing with other kinds of high-radix topologies.(4)We present xNetSim, an flit-based simulator for larger scale high performance interconnection network based on OMNeT++, which consists of the topology module,the packet module and the router module. As to meet the need of performance evaluation,we adjust the output of topology and routing computation for different topologies.Based on the analysis and comparison of theoretical analysis and the interconnection network Simulator xNetSim, the results of evaluation show that the proposed topology SuperStar with better scalability and low network latency.