基于离散萤火虫的众核SoC映射方法研究

发布时间：2018-12-24 08:03

【摘要】：NoC映射问题是NoC领域的研究热点之一,映射的好坏直接影响片上网络系统整体的延迟、功耗等方面的性能。现有研究大多数都是应用启发式优化算法来解决片上网络映射问题,但算法优化时间过长,当众核规模大时,需要占用大量的计算时间和资源,而改进的算法以牺牲解的精度提升优化速度。萤火虫算法是一种新型的启发式优化算法,本文基于离散萤火虫算法来解决片上网络映射问题,主要工作如下：1、基于离散萤火虫算法的2D片上网络映射算法。针对现有的离散萤火虫算法不能解决2D NoC映射问题,重新定义萤火虫类,使离散萤火虫算法与片上网络映射联系起来。重新定义2D映射算法初始化规则,随机给出初始化映射方案。重新定义萤火虫之间的距离的计算方法,作为判断移动的影响因子。定义萤火虫移动规则分为α移动步骤和β移动步骤,用于寻找更优的映射方案。2D片上网络映射实验结果表明,与初始化中的最优映射方案相比,全局最优映射方案的总通讯损耗优化了41.46%。与离散粒子群映射算法相比,探索能力更强,平均微片延迟最大优化了58.005%,平均包延迟最大优化了42.874%。2、基于离散萤火虫算法的3D片上网络映射算法：针对3D NoC拓扑结构特点,重新定义初始化规则、定义萤火虫之间距离的计算方法。3D片上网络映射实验结果表明,与基于离散粒子群映射算法相比,可找到最优映射方案。与离散粒子群映射算法相比,稳定性更好,所用的CPU执行时间最大优化了69.656%。3、设计一种支持多路接入的网络接口和一种基于自相似的流量发生器：针对特征应用中子任务一对多、多对一通讯关系,提出一种支持多路接入的网络接口RTL级模型,通过优先级列表来轮转仲裁请求的流量发生器,与已有片上网络仿真平台组合完成映射仿真实验。为了更好的模拟实际网络中的流量特性,设计一种基于自相似的流量发生器,使用ON/OFF模型产生发包周期和空闲周期长度,根据此长度进行发包。多路接入的网络接口和基于自相似的流量发生器仿真实验表明,两个模块功能是正确的。
[Abstract]:The problem of NoC mapping is one of the hotspots in the field of NoC. The quality of the mapping directly affects the performance of the whole network system on chip in terms of delay, power consumption and so on. Most of the existing researches apply heuristic optimization algorithm to solve the problem of on-chip network mapping, but the optimization time of the algorithm is too long, when the public kernel scale is large, it needs a lot of computing time and resources. The improved algorithm improves the optimization speed by sacrificing the accuracy of the solution. Firefly algorithm is a new heuristic optimization algorithm. This paper solves the problem of network mapping on chip based on discrete firefly algorithm. The main work is as follows: 1. 2D on-chip network mapping algorithm based on discrete firefly algorithm. Because the existing discrete fireflies algorithm can not solve the 2D NoC mapping problem, the class of fireflies is redefined so that the discrete fireflies algorithm is associated with the on-chip network mapping. The initialization rules of 2D mapping algorithm are redefined, and the initialization mapping scheme is given at random. The calculation method of redefining the distance between fireflies is used as a factor to judge movement. Firefly moving rules are defined as 伪 moving step and 尾 moving step to find a better mapping scheme. The experimental results of 2D on-chip network mapping show that, compared with the optimal mapping scheme in initialization, The total communication loss of the global optimal mapping scheme is optimized by 41.46. Compared with discrete Particle Swarm Optimization (DPSO), the algorithm has more searching ability. The maximum average microchip delay is 58.005, and the average packet delay is 42.874.2. 3D On-chip Network Mapping algorithm based on discrete Firefly algorithm: redefining initialization rules and calculating distance between fireflies according to 3D NoC topology characteristics. The experimental results of 3D on-chip network mapping show that, Compared with the discrete particle swarm optimization algorithm, the optimal mapping scheme can be found. Compared with the discrete Particle Swarm Optimization (DPSO) algorithm, the stability of the algorithm is better. The maximum execution time of CPU is optimized by 69.656. A network interface supporting multiple access and a flow generator based on self-similarity are designed. According to the one-to-many, many-to-one communication relationship of feature application neutron tasks, a network interface RTL level model is proposed to support multiplex access. The flow generator of arbitration request is rotated by priority list, and the mapping simulation experiment is completed by combining with the existing on-chip network simulation platform. In order to better simulate the traffic characteristics in the real network, a self-similar flow generator is designed. The ON/OFF model is used to generate the length of the outgoing cycle and the idle period, according to which the contract is sent. The network interface of multiple access and the simulation experiment of flow generator based on self-similarity show that the function of the two modules is correct.
【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】