张毅, 何卫锋. 基于Fat-tree的高性能互联网络性能优化与分析[J]. 微电子学与计算机, 2018, 35(8): 36-41.
引用本文: 张毅, 何卫锋. 基于Fat-tree的高性能互联网络性能优化与分析[J]. 微电子学与计算机, 2018, 35(8): 36-41.
ZHANG Yi, HE Wei-feng. Optimization and Evaluation of Fat-tree Based High Performance Interconnect Network[J]. Microelectronics & Computer, 2018, 35(8): 36-41.
Citation: ZHANG Yi, HE Wei-feng. Optimization and Evaluation of Fat-tree Based High Performance Interconnect Network[J]. Microelectronics & Computer, 2018, 35(8): 36-41.

基于Fat-tree的高性能互联网络性能优化与分析

Optimization and Evaluation of Fat-tree Based High Performance Interconnect Network

  • 摘要: 高性能互联网络对于高性能计算机的性能具有重要的影响, 而高性能网络中各个模块的设计都会影响到网络的最终性能, 因而前期的网络仿真显得尤为重要.HPCSim是基于OMNeT++平台自主开发的高性能网络仿真器, 包括了网络拓扑和仿真参数自动生成工具PySim以及仿真结果分析和可视化工具PyVisual.仿真器中对全缓存高阶路由器结构进行了改进, 提出了一种5级流水线的高阶路由器, 能够减少路由信息缓存开销同时将网络吞吐率从基准路由器下的0.75提升至0.95.此外, 还对硅光互联链路进行了逻辑抽象, 并在Fat-tree拓扑上对链路性能和网络吞吐率的关系进行了分析.

     

    Abstract: High performance interconnect network has a significant influence on the performance of high performance computer. As the difference in network model design will result in different network performance, the early network simulation is needed. HPCSim is an independently developed high performance interconnect network simulator based on OMNeT++. It includes topology and parameter file generator named PySim, and simulation data analyzer and visualizer named PyVisual. An improved fully-buffered router structure is proposed, which achieves 5 stage pipeline, and reduces routing computation states buffer as well as promotes throughput from 0.75(baseline router) to 0.95. Moreover, a link model is inferred from photonic network, and the link performance simulation is run based on Fat-tree network to analyze the relation between link bandwidth and network throughput.

     

/

返回文章
返回