【作者】 王晓锋；

【导师】 方滨兴；

【作者基本信息】 哈尔滨工业大学 ， 计算机系统结构， 2007， 博士

【摘要】 网络模拟是研究网络行为、评价网络协议的重要手段,具有重要的学术研究与应用价值。随着计算机网络的快速发展,所需模拟的网络规模越来越大,而大规模离散事件网络模拟所需的大量计算开销无法让人接受。如何减少大规模网络模拟的运行时间、提高模拟性能,使得能快速模拟十万、百万节点级规模的网络,甚至使得模拟整个互联网成为可能,是一个亟待研究的问题,因此本文致力于此。本文以大规模离散事件网络模拟技术为研究对象,分析评价影响网络模拟性能的各种因素,研究提高大规模离散事件网络模拟性能的方法。本文研究的主要内容包括如下五个部分:第一,为研究如何提高网络模拟性能,分析了影响并行网络模拟性能的各种因素,包括负载平衡、通信开销、安全模拟时间窗口长度、模拟运行环境、模拟应用等,并建立了关于这些因素的并行网络模拟性能估计模型。与传统的估计模型相比,该估计模型能够为如何提高网络模拟性能提供指导。通过对各种因素的分析以及估计模型的建立,明确了提高网络模拟性能的具体思路,为后续章节的研究提供基础。第二,鉴于当前基于拓扑图划分工具的并行网络模拟划分方法存在易陷入局部最优以及不能实现合理划分的不足,提出了并行网络模拟拓扑的优化划分方法。该优化划分方法采用改进的模拟退火算法实现对所要模拟的网络拓扑的划分,所使用的目标函数为并行网络模拟性能估计模型。基于模拟退火算法的划分不易陷入局部极值;基于估计模型的划分能够综合考虑各种影响网络模拟性能的因素,实现对模拟任务的合理划分,因此该方法能有效提高并行网络模拟的性能。第三,鉴于大规模离散事件网络模拟需要极多的计算开销以描述大量的数据包转发,提出了基于动态连续计算的快速离散事件网络模拟方法。该方法将数据包转发模拟采用连续计算的方法描述,减少了传统网络模拟过程中数据包转发模拟的不连续性带来的开销;分析了连续计算带来的不真实性,并为保证连续计算具有一定真实性,提出了动态连续计算方法,即对数据包转发动态有选择的连续计算。第四,为减少流量生成器产生的背景流量数据包的模拟开销,提出了基于数据包采样的背景流量简约模拟方法。在背景流量生成时,该方法采用“采样”技术只产生少量的背景流量数据包;在拥塞链路上,依靠这些少量数据包“恢复”出原始背景流量的所有数据包;依靠恢复得到的数据包信息,实现对路由器缓冲队列以及前景流量的模拟。实验证明,该方法在提高背景流量模拟性能的同时,保证模拟结果的真实性。第五,基于上述研究内容,并结合网络模拟拓扑生成模块、远程路由配置模块、模拟应用描述脚本生成模块等处理模块,建立了并行网络模拟应用平台,该平台具有性能高、可用性强等特点。最后,通过一个具体的大规模并行网络模拟实例,说明该平台的使用步骤。更多还原

【Abstract】 Simulation is a key method to do research on network behavior and to analyze network protocols, and it is of great value in science research and application. With the rapid development of computer networks, the scale of network for simulation grows larger and larger, yet the computation overhead of the large-scale discrete event network simulation can hardly be satisfied. How to reduce the running time of large-scale network simulation, that is, how to improve the simulation performance, so as to make it possible to simulate a network with a scale of million-node fastly, or even to simulate the Internet, is a most challenging problem. So, this paper is dedicated to this problem.Taking large-scale discrete event network simulation as the research object, this dissertation analyzes various factors that may affect the performance of network simulation, and investigates solutions to improve the performance of network simulation. This paper is mainly composed of the following five parts:First, to investigate how to improve performance of network simulation, various factors, including load balancing, communication overhead, lookahead, simulation running platform, and the application to be simulated etc., are analyzed, and based on these factors, a model for estimating the performance of parallel network simulation is developed. Compared to the traditional estimating models, this model can give guidance to how to improve the performance of network simulation. Through analysis of those various factors and establishment of the estimating model, the actual methods on how to improve the performance of network simulation are put forward, which are the basis of the following research work.Second, as there are shortcomings, such as convergence to local minima and unreasonable patitioning result, of the current parallel network simulation partitioning methods which are based on topology graph partitioning tools, a method for optimized partitioning for parallel network simulation topology is developed. In this optimized partitioning method, the network topology for simulation is partitioned by the improved simulated annealing, and the model for estimating the performance of parallel network simulation is treated as the object function. Depending on the simulated annealing, the partitioning result does not tend to converge to local minima, and depending on the estimating model, the partitioning method can take various factors that may affect the performance of network simulation into consideration, and partition the simulation task reasonably, so this method can improve the performance of parallel network simulation efficiently.Third, as during the procedure of large-scale discrete event network simulation, much computation overhead is needed to deal with a great lot of packet transmissions, a fast discrete event network simulation method based on dynamic continuous computing is put forward. In this method, the simulation of packet transmission is depicted by continuous computing in order to reduce the overhead generated by the discontinuity of the simulation of packet transmission in the traditional network simulation. The inaacuracy of the continuous computing is analysed, and to keep the accuracy, the dynamic continuous computing method, which means to process the packet transmissions with continuous computing dynamically and selectively, is put forward.Fourth, to reduce the simulation overhead of background traffic which is generated by the traffic generator, a simplified background traffic simulation method based on packet sampling is developed. When generating the background traffic, the method only generate part of the packets in the original background traffic by sampling; In the congested links, all the packets in the original background traffic are recovered according to these few packets; Depending on these recoverd packets, the buffers of the routers and the foreground traffics are simulated. Expriments show that this method can improve the performance of background traffic simulation, while keeping the accuracy of simulation result.Fifth, combining the topology generating module of network simulation, the configuration module of remote route, the script generating module of simulation application, and the studies metioned above, a parallel network simulation application platform is developed. The application platform is of high performance and convenience. At last, the procedure of how to use the platform is demonstrated through an example of large-scale parallel network simulation.更多还原