【作者】 王坚；

【导师】 彭启琮；

【作者基本信息】 电子科技大学 ， 通信与信息系统， 2011， 博士

【摘要】 片上网络技术已成为目前国内外研究的热点,它对于解决片上多处理器之间的通信瓶颈,提高片上多处理器系统的性能,促进高性能计算机的发展有着重要的意义。为了提高片上网络的研发效率,减少片上网络的开发周期和开发成本,研究一种能快速准确地评估片上网络性能的方法是有意义的。此外,由于片上通信环境的特殊性,片上网络对系统资源、面积和功耗等开销有严格的限制,因此,研究在约束条件下片上网络的性能优化方法是很必要的。本论文正是针对上述问题,在以下几个方面对片上网络的通信性能分析与优化问题进行了探索性研究:1.从分析片上网络的系统特性和应用特性入手,解决片上网络通信性能的分析建模问题;对采用虫洞机制的片上网络,本文利用半马尔科夫过程描述了路由节点的各种状态,并以此为基础分析虫洞机制路由节点的性能和片上网络的通信性能。对采用存储转发/虚切片交换机制的片上网络,本文将路由节点中的各个输入缓存抽象为排队系统,通过对排队系统的分析求解数据通过路由节点的延迟,并在此基础上分析整个片上网络的通信性能。2.在片上网络缓存资源的约束下,提出新的片上网络缓存资源优化方法,以提高片上网络通信性能;对片上网络路由节点中的缓存资源,本文研究了缓存优化配置的方法,以在不增加缓存总开销的情况下提高片上网络通信性能;对片上网络接口中的缓存资源,本文研究了最佳的网络接口缓存大小,以在保证数据服务质量的同时避免缓存浪费。3.在片上网络带宽资源的约束下,研究了链路带宽对片上网络性能的影响,并对非均匀链路带宽的片上网络进行了仿真以修正片上网络的通信性能分析模型。在此基础上,本文提出了片上网络链路带宽的优化设计算法,以降低系统对带宽的需求,从而为片上网络面积和功耗优化提供了理论依据。4.研究了特定应用在不同映射方案下对片上资源的需求,并基于蚁群优化理论提出了约束条件下的片上网络映射算法以降低片上网络的开销、提高片上网络的性能;然后分别在延迟约束、缓存资源约束和带宽资源约束的前提下验证了本文映射算法的可行性。5.对片上网络常见调度策略进行了总结,并分析了他们的不足。在此基础上,结合片上网络自身的特点,研究了片上网络中的动态调度策略,并提出了相应的调度器设计方案,以解决片上网络调度中的“饿死”问题。同以往的调度算法相比,本文的动态仲裁路由器对片上网络通信性能有一定改善,并减少了网络接口处对缓存的需求。更多还原

【Abstract】 NoC (Network-on-Chip) has been focused by a lot of researchers for years. The NoC solution is used to solve the bottleneck of the on chip communication between processors, to improve the performance of overall system, and to promote the development of high performance computer. In order to improve the efficiency of NoC design and reduce the time-to-market of NoC product, it is important to develop a useful tool to evaluate the NoC performance. Moreover, since the NoC is limited by the on-chip resources, area and power consumption, it is crucial to research the optimization method with various constraint conditions to improve the NoC performance.In this thesis, we make innovative researches on NoC and the main results are shown as follow:1. By analyzing the system characters and the application characters, we propose a modeling method for NoC communication performance. For the wormhole NoCs, a semi-Markov process is developed to describe the work state of router. Then, we analyze the router performance and NoC communication performance based on this semi-Markov process. For the store-and-forward/virtual-cut-through NoCs, each router buffer is abstracted as a queuing system. By modeling the NoCs in this queuing manner, we can calculate the average packet latency of each router and then analyze the NoC communication performance.2. With respect to the constraint condition of buffer resources, a method is proposed to optimize the NoC buffer resources, such that the NoC communication performance is optimized. For the buffer optimization of NoC router, a novel buffer allocation method is proposed to maximize the NoC performance without any additional buffer cost. For the optimization of the buffer of network interface, a novel buffer sizing method is proposed to determine the proper buffer size of network interface, which maintains the data Quality-of-Service while avoids the waste of buffer resources.3. With respect to the constraint condition of bandwidth, we study the influence of bandwidth on the NoC performance, simulate the performance of NoC with different link bandwidth and modify the analytical model for NoC performance. Then, a novel bandwidth optimization method is proposed to determine the bandwidth for all links in NoC, such that the cost of NoC is minimized. Our model also provides the theory basis for the optimization of NoC area and power consumption.4. By analyzing the different resources requirements in different NoC mapping, we propose an ACO (Ant Colony Optimization) based mapping algorithm to minimize the cost of NoC and improve the performance of NoC. Then, we evaluate the effectiveness of our algorithm under the constraints of latency, buffer resources and bandwidth, respectively.5. After analyzing the disadvantages of the common used schedule strategies in NoC, we propose a dynamic schedule strategy and design the corresponding arbiter to implement the dynamic schedule strategy, such that the‘starvation’problem in NoC schedule is solved. Compared with other schedule strategies, our dynamic schedule strategy can improve the NoC performance and minimize the cost of buffer resources in network interface.更多还原