基于网络演算的片上网络积压上界优化方法研究

发布时间：2018-09-04 18:40

【摘要】：片上网络(Network-on-Chip,NoC)最差情形的分析不仅影响集成电路性能,而且影响芯片的可靠性与稳定性,因而成为NoC研究领域的热点之一。最差情形性能包括延迟上界、积压上界等指标。现有工作大多集中在NoC延迟上界的研究,对于积压上界计算分析的工作相对较少,需要在积压上界的计算分析、积压上界优化等方面深入研究。本文以网络演算为基础,结合萤火虫算法,对于网络的积压上界进行计算和优化。主要工作如下:(1)建立片上网络积压上界计算模型以网络演算为基础,针对于片上网络的积压上界,建立一种计算片上网络全局积压和单点积压的模型。通过计算单条流的积压,将其分配到该条流经过的每个节点,选取所有经过当前节点的业务流产生的最高积压作为节点积压上界,建立积压矩阵,计算全局网络积压上界。(2)建立针对单节点积压上界进行优化算法针对单个节点积压上界过大造成的网络节点拥塞、网络延迟增高等问题,利用(1)中方法,得到单个节点积压上界。将此积压上界作为目标函数,结合萤火虫算法,利用五种特征应用VOPD、MWD、263ENC MP3DEC、MP3ENC MP3DEC、263DEC MP3DEC,选取不同的网络性能参数,对积压上界进行优化。实验结果表明,与优化前相比,在不同的发包速率、路由节点转发速率、路由节点转发延迟对于节点积压上界的优化分别达到30.26%、51.04%、30.31%。(3)建立针对全局积压上界进行优化算法针对全局网络积压上界对网络造成的面积冗余、资源开销等问题,利用(1)中方法,得到全局积压上界。以此为目标函数,利用特征应用VOPD,选取不同的网络性能参数,通过萤火虫算法,对全局积压上界进行优化。实验表明,与优化前相比,不同的发包速率、路由节点转发速率、路由节点转发延迟等因素对于网络全局积压的优化都在2-4%之间。
[Abstract]:The analysis of the worst case of on-chip network (Network-on-Chip,NoC) not only affects the performance of IC, but also affects the reliability and stability of the chip, so it has become one of the hotspots in the field of NoC research. The worst performance includes delay upper bound, backlog upper bound and so on. Most of the existing work focuses on the research of the NoC delay upper bound, but there is relatively little work on the backlog upper bound calculation and analysis, which needs to be further studied in such aspects as the calculation analysis of the backlog upper bound, the optimization of the backlog upper bound, and so on. Based on network calculus and firefly algorithm, this paper computes and optimizes the upper bound of network backlog. The main works are as follows: (1) A model of computing the global and single-point network backlog on a chip is established based on the network calculus. By calculating the backlog of a single stream and assigning it to each node passing through, the maximum backlog generated by all the traffic flows through the current node is selected as the upper bound of the backlog, and the backlog matrix is established. The global network backlog upper bound is calculated. (2) the optimization algorithm for the single node backlog upper bound is established to solve the problems of network node congestion and network delay increase caused by the oversize of the single node backlog upper bound, and the method in (1) is used. The upper bound of single node backlog is obtained. The upper bound of the backlog is taken as the objective function and the firefly algorithm is used to optimize the upper bound of the backlog by using VOPD,MWD,263ENC MP3DEC,MP3ENC MP3DEC,263DEC MP3DEC, to select different network performance parameters. The experimental results show that the routing node forwarding rate is different from that before the optimization. The routing node forwarding delay optimizes the upper bound of the node backlog to 30.26 / 51.04 / 30.31 respectively. (3) an optimization algorithm for the global backlog upper bound is established to solve the problems of the area redundancy and resource overhead caused by the global network backlog upper bound, using (1) the method in the paper. The upper bound of global backlog is obtained. Taking this as the objective function, using the feature of VOPD, to select different network performance parameters, the global backlog upper bound is optimized by the firefly algorithm. The experimental results show that the optimization of network global backlog is between 2-4% due to different packet delivery rate and routing node forwarding delay.
【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】