异构多核片上网络布局优化研究与实现

发布时间：2018-06-12 09:52

本文选题：片上网络 + 异构多核　；参考：《合肥工业大学》2017年硕士论文

【摘要】：随着集成电路集成度不断提高,特征尺寸不断减小,工作频率不断上升,传统总线结构的多功能片上系统(SoC)的扩展性差、并行度低以及全局时钟同步困难等缺点逐渐显露出来。为了克服总线型SoC的固有缺点,片上网络(NoC)的概念被提出。NoC从通信架构层面有效地解决了总线型SoC扩展性差和并行度低等问题,并以其优良的特性成为研究的热点。片上网络的布局决定着网络中各节点在网络中的分布,影响着资源节点间的距离、通信流量及热量等特征,进而对系统的功耗、面积、时延、QoS等性能产生重大影响。NoC的布局优化以提升NoC系统的性能为目标,包括NoC体系结构的优化和基于NoC应用的优化。NoC体系结构的优化包括构建拓扑结构,设计路由算法及交换机制等方面。基于NoC应用的布局优化实现应用任务在IP核中的合理分配和映射,同时根据性能要求确定各IP核在网络中的合理布局,以满足不同性能的需求。本文主要研究了基于NoC应用的布局优化,结合目标系统,实现了基于NoC的异构多核的多目标布局优化。本文的主要工作如下:首先深入研究了一种面向多应用NoC系统的IPCore分配方法,并根据该方法提取了目标系统的通信特征图。根据布局中的映射阶段中的需求和约束条件,建立了目标系统的多目标优化模型,通过平衡各优化目标间的需求,得出目标系统多目标优化的目标函数。针对现有布局评估方案的缺点,提出了新的评估方案,使用智能优化算法完成了异构多核系统的布局优化。其次,根据本文所提出的评估方案,使用C++语言搭建了一个布局优化平台,该平台采用层次化的设计方法,包括优化算法层、任务层、控制层、网络层和统计层,详细的介绍了每个层次的设计实现方案及其工作流程。布局优化平台通过优化算法层控制整个随机搜索过程,根据通信特征图灵活的构建各种通信任务,同时模拟任务在系统中的执行过程可以对各布局方案进行准确的评估。最后,在布局优化后的系统中进行不同规模的矩阵计算实验,并与随机布局系统进行对比,以评估布局优化后的系统性能。实验结果表明:布局优化后的系统中算法执行效率最大提升了 36.88%,同时具有较低的通信功耗及延时。
[Abstract]:With the increasing integration of integrated circuits, the feature size is decreasing, the working frequency is increasing, the expansibility of SoC-based multifunctional system with traditional bus architecture is poor, the parallelism is low, and the global clock synchronization is difficult, and so on. In order to overcome the inherent shortcomings of bus-type SoC, the concept of on-chip network (NOC) has been proposed to effectively solve the problems of poor scalability and low parallelism of bus-based SoC from the level of communication architecture, and has become a hot research topic with its excellent characteristics. The layout of the on-chip network determines the distribution of the nodes in the network, affects the distance, communication flow and heat between the resource nodes, and then the power consumption and area of the system. The layout optimization of NOC has a significant impact on performance, such as delay, QoS, etc., which aims to improve the performance of NoC system, including the optimization of NoC architecture and the optimization of NoC architecture based on NoC application, including the construction of topology. The routing algorithm and switching mechanism are designed. The layout optimization based on NOC can realize the rational allocation and mapping of application tasks in IP core, and determine the reasonable layout of each IP core in the network according to the performance requirements, so as to meet the requirements of different performance. In this paper, the layout optimization based on NoC application is studied, and the multi-objective layout optimization of heterogeneous multi-core based on NoC is realized in combination with the target system. The main work of this paper is as follows: firstly, an IPCore allocation method for multi-application Noc system is studied, and the communication feature diagram of the target system is extracted according to this method. According to the requirements and constraints in the mapping stage in the layout, the multi-objective optimization model of the target system is established, and the objective function of the multi-objective optimization of the target system is obtained by balancing the requirements among the various optimization objectives. In view of the shortcomings of the existing layout evaluation schemes, a new evaluation scheme is proposed, and the layout optimization of heterogeneous multi-core systems is accomplished by using intelligent optimization algorithm. Secondly, according to the evaluation scheme proposed in this paper, a layout optimization platform is built with C language. The platform adopts hierarchical design methods, including optimization algorithm layer, task layer, control layer, network layer and statistical layer. The design and implementation of each level and its workflow are introduced in detail. The layout optimization platform controls the whole random search process through the optimization algorithm layer, constructs various communication tasks flexibly according to the communication characteristic graph, and simulates the execution process of the task in the system, which can accurately evaluate each layout scheme. Finally, the matrix calculation experiments of different scales are carried out in the optimized layout system, and compared with the random layout system to evaluate the performance of the optimized layout system. The experimental results show that the efficiency of the algorithm increases by 36.88, and the communication power consumption and delay are lower.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN47

【相似文献】