多路径最短路由NoC的网络演算模型及性能优化

发布时间：2018-01-12 07:38

本文关键词：多路径最短路由NoC的网络演算模型及性能优化　出处：《合肥工业大学》2015年硕士论文　论文类型：学位论文

【摘要】：片上网络技术(NoC)日趋成熟,并己开始逐步应用于产业界。NoC最差情形性能评估是NoC设计的重要指标,然而对其进行的研究工作相比对NoC平均性能的研究来说却较少。了解NoC的最差情形性能有利于设计者清楚片上网络的性能边界,增强设计信心。本文以网络演算为基础,以多路径最短路由NoC中最差情形延迟上界为研究对象,建立最差情形性能下的延迟上界分析模型,构建片上冲突预测机制,并提出优化策略对最差情形延迟上界进行优化。主要工作如下：(1)多出口聚集流分析模型针对现有Fidler分析模型对网络最差情形延迟上界计算不够准确的问题,利用单出口聚集流分析模型在求解等价服务曲线时处理自由参量的特点,本文构建基于网络演算的多出口聚集流分析模型,使得求解过程更加方便,求解结果同仿真结果之间的差值相对Fidler模型平均优化达72.2%。(2)描述片上网络冲突情况针对片上网络冲突情况复杂问题,本文构建冲突矩阵来描述网络中业务流之间的冲突信息,该冲突矩阵中元素包含冲突业务流的流量特性以及其拆分情况。冲突矩阵可以清楚的表明目标业务流在各个路由节点的冲突情况。在清楚网络冲突情况后,可以得到目标业务流每条子路径上的冲突大小。(3)电阻非均匀拆分策略针对多路径路由片上网络拥塞严重的问题,本文提出了电阻非均匀拆分策略。首先使用冲突矩阵得到各子业务流路径冲突系数大小,描述路径冲突大小的冲突系数可以看作该路径的“电阻”,然后类比并联电路电流分配方法,提出电阻非均匀拆分策略来进行目标业务流的流量分配。以此来将大部分数据流发往冲突较小的路径,小部分数据流发往冲突较大的路径,从而均衡负载和优化最差情形延迟上界。采用该策略相对均匀拆分的多路径路由片上网络在延迟上界方面有平均5%的优化。(4)基于Verilog的RTL级仿真平台针对现有仿真平台无法实现电阻非均匀拆分策略的问题,本文搭建了基于Verilog的RTL级仿真平台,该仿真平台将电阻非均匀拆分模块封装在一个离线控制核中,并使用该核所得业务流拆分比来进行路由节点的业务流流量分配,以此优化网络最差情形性能。
[Abstract]:Network on chip (NoC) technology has become more mature, and has been gradually applied in the industry.NoC the worst case performance evaluation is an important index of NoC design, but the research work on the research of NoC compared with the average performance is less. NoC know the worst case performance is beneficial to clear performance bounds on chip designers the network design, enhance confidence. Based on network calculus, with the shortest path routing NoC in the worst case delay bound as the research object, establish the analysis model of delay bound in the worst case performance of the construction on conflict prediction mechanism, and puts forward the optimization of the worst case delay bound optimization strategy. The work is as follows: (1) analysis model based on Fidler analysis model of network delay bound in the worst case calculation inaccuracy problem of flow aggregation and multi outlet, use analysis model in solving flow aggregation of single export The characteristics of the equivalent service curve when dealing with free parameters, this paper constructs the flow analysis model of multi outlet aggregation based on network calculus, which makes the calculation process more convenient, the difference between the results obtained with the simulation results of the relative Fidler model optimal average reached 72.2%. (2) to describe the situation of internet tablet for network on chip network conflict conflict complex problems, this paper constructs the conflict matrix to describe the conflict between information network traffic, elements of the conflict matrix contains flow characteristics of traffic flow and the resolution of the conflict. The conflict matrix can be clearly marked in the conflict table Mingmu business flow of each routing node in the network. Clear the situation after the conflict, to get traffic conflict size per note on the path. (3) uniform resolution strategy for multi path routing on network congestion problem of non resistance, non uniform resistance is proposed in this paper. Split strategy. First use the conflict matrix to get the business flow path conflict coefficient, describe the path conflict size conflict coefficient can be regarded as the path of "resistance", then the current distribution of parallel circuit analogy method, the resistance of non uniform flow distribution resolution strategies to target business flow. In order to most of the data flow path to the conflict of small, small part of the data stream sent to the larger conflict path, so as to balance the load and optimize the worst case delay bound. Multi path routing with relatively uniform resolution on the strategy of network has an average of 5% in the delay bound optimization. (4) based on RTL Verilog simulation platform for the existing simulation the platform can not achieve the resistance of non-uniform split strategy, this paper built a RTL based on the Verilog simulation platform, the simulation platform will resistance non-uniform resolution module encapsulation In an off-line control core, the service flow split ratio of the core is used to allocate the traffic flow of the routing node, so as to optimize the worst case performance of the network.

【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN47

【相似文献】