【作者】 徐锐；

【导师】 余翔湛；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2008， 硕士

【摘要】 在网络模拟研究中,单节点计算环境无法模拟大规模网络。并行网络模拟通过引入更多的计算节点可以解决这个问题。方法是将模拟拓扑划分为几个部分,分别由各个计算节点负责模拟,可以实现模拟大规模网络并能减少模拟时间。现有的网络模拟拓扑划分通常需要计算环境的指导,即把计算环境和模拟拓扑作为划分算法的输入,在划分中考虑到节点的计算性能,并尽量减少各个划分块之间的链路负载,以期望能降低并行模拟的通信开销。现有的划分算法,在计算环境规模较小的情况下,能根据计算节点的性能,合理的把模拟拓扑映射到计算环境中,能收到较好的划分效果。但是,对于大规模的并行计算环境,计算节点数目众多,性能层次不齐,无法确定为一个特定模拟任务分配资源的多少。因此,现有划分算法受到计算环境的限制,不再适用于大规模计算平台的网路模拟任务划分。本文提出了一种基于模拟拓扑特征的并行网络模拟的拓扑划分方法,该划分方案通过分析基准实验,引入了影响并行模拟的几个主要因素,如负载均衡和减少通信开销,最终能实现拓扑划分中,通信开销所占比例较小,并利用大规模并行计算网络中节点性能的多样性,实现了负载均衡,可以提高并行网络模拟平台的吞吐率和模拟效率。较之现有的划分方法,不需要计算环境的指导,完全从分析模拟拓扑本身出发,提高了并行模拟加速比,具有很强的应用价值。最后,本文还整合了现有的划分工具,设计了大规模网络模拟平台的管理调度系统,方便管理和调度计算资源,可以实现了一键式并行网络模拟,较之传统的手工方式,极大的方便了研究人员使用平台,提高了并行网络模拟的易用性。更多还原

【Abstract】 In the network simulation research, large-scale network can not be simulated in single-node due to limited computing capacity. With more computing nodes, Parallel network simulation can solve this problem. The key method is dividing the network into several parts of topology, and then calculating by multi-nodes, which can expand the scale of the simulation topology and reduce the simulation time.The existing topology partition method, using both the computing environment and the simulation topology as input, considers the difference of node’s calculating performance, and maps the simulation topology to computing environment with loading balance and minimizing the communication overhead. This method does a good job in small-scale computing environment, but in large-scale parallel network simulation platform, because of plentiful nodes and variety of the computing performance, this method can not make sure how much resource should be allocated for a simulation topology. Because of needing guidance of computing environment, existing partition method can not work efficiently for large-scale parallel network simulation platform.This paper designs a new partition method based on the simulation topology itself, with minimizing the communication overhead. For the variety of the computing performance of the large-scale platform, loading balance can be easily got. First, we consider the factor of parallel network simulation and do an experiment to analyze it, and then design the partition algorithm; finally we use some case to verify the performance. Comparing the existing method, this one does not need the guidance of the computing environment, and can discover the maximum speedup in a higher possibility, which can improve the efficiency and throughput of platform.At last, we design a system for a large-scale network simulation platform to facilitate the scheduling and management of computing resources and to achieve a one-button-simulation, greatly helping the researchers to use platform conveniently.更多还原