面向高密度计算的NoC平台多发射技术研究

发布时间：2018-04-17 08:28

本文选题：高密度计算 + 片上网络　；参考：《合肥工业大学》2015年硕士论文

【摘要】：随着科技的发展和生产生活的信息化,诸多领域对处理器的计算能力提出了更高的诉求。在算法方面,人们提出了多种高效的算法来解决实际问题,其中一部分的发展趋势是计算密度急剧提高和支持并行计算。而在计算平台方面,也演变出了各种经过特别优化的专用平台用于满足严苛的需求。基于片上网络(NoC)架构的异构多核处理器是其中一个重要的分支。研究提高异构多核NoC处理器的性能,是半导体行业一个重要课题。本文对上述问题,对面向高密度计算的NoC平台上多发射技术进行了研究和实验。本文所做主要工作如下：一、针对单发射原型系统NoC平台进行多发射改进分析,提出了基于可扩展发射通道的主控制器设计,管理多发射的动态调度。针对多发射系统运算簇资源使用特性,设计了用于运算簇资源动态分配的空闲簇队列和相应排序算法。针对计算子任务的动态访存特性设计了地址映射关系表,用于动态映射虚拟地址/实地址。针对动态分配运算簇和实地址的特性,设计了配置信息填充模块。针对多发射改进控制特性,进行了原型系统的DDR去通道化设计和状态网络功能增强设计,用于适应多发射改进。针对瓶颈子任务,提出了支持局部算法级乱序多发射的优化设计。二、对本文设计的多发射主控制器建立了软件功能模型,并加入了时间变量用于获取系统运行的性能信息。在实验准备阶段,完成对一种高密度算法的分析与任务映射。随后,使用软件模型加载高密度算法对系统进行了多发射功能验证和性能提升验证,并对比了是否开启局部算法级乱序多发射功能对系统性能提升的影响。根据性能分析结果,提出了优化系统运算簇构成和算法映射的方法,并经软件模型验证有效。最后,根据奇数个算法前/后端计算与偶数个的差距提出了解决办法,经软件模型验证有效。三、对本文设计的多发射主控制器进行了RTL级代码编写,并进行了系统硬件集成。首先使用ModelSim仿真验证了子任务规模4096：1缩比下的功能正确性,随后在FPGA开发板上完成全规模实验,验证了设计的可行性和软件模型获取性能信息的准确性。本文改进设计的多发射系统具有如下特点：(1)任务调度方式适用于高计算密度类型的算法,多发射方式适用于在高密度算法中进行多路并行计算。(2)支持各发射通道进行不同类型计算。(3)多发射对象为计算子任务,采用动态调度方式。(4)多发射通道共享系统资源,通道数量可根据需要进行扩展。(5)根据不同发射通道动态分配最优运算簇资源。(6)精简配置信息长度,动态映射实地址/虚拟地址空间。(7)具有可选择开启的局部乱序多发射优化功能。
[Abstract]:With the development of science and technology and the production of life information, put forward higher demands on the processor computing power in many fields. In the algorithm, it proposes several efficient algorithms to solve practical problems, which is part of the development trend of computing density increases sharply and parallel computing. In computing platform, but also the evolution of after a variety of special platform specially optimized to meet stringent requirements. Based on network on chip (NoC) architecture of heterogeneous multi-core processor is one of the important branches of research. To improve the performance of heterogeneous multi-core NoC processor, is an important issue in the semiconductor industry. Based on the above problems, opposite to high density calculation the NoC platform on the launch technology for research and experiment. The main work of this paper is as follows: first, for the single launch prototype system of NoC platform for multi emission improvement analysis, put forward The design of the main controller can extend the transmission channel based on dynamic scheduling and management of multiple transmissions. For many launch system cluster resource usage characteristics for free cluster queue cluster dynamic resource allocation and the corresponding sorting algorithm is designed. According to the dynamic calculation of sub task memory address mapping table design characteristics, for dynamic mapping virtual address / real address. According to the dynamic allocation of cluster and real address characteristics, design the configuration information filling module. For many improved emission control characteristics, the prototype system of DDR to channel design and network function design enhancement, to adapt to the multi launch. Aiming at the bottle neck improvement task, put forward the optimum design support local algorithm level reordering multiple transmit. Two, the multi transmitting main controller to establish software function model, and add the time variable for acquiring system The performance of information operation. In preparation for the experiment, analysis and mapping of tasks to complete a high density algorithm. Then, using the software model loading high density algorithm of multiple transmit function verification and performance improvement verification and comparison of whether to open the local algorithm level reordering multiple transmit function of system performance improvement according to the results of the performance analysis, the method of optimizing the system cluster structure and mapping algorithm, and the software model is validated. Finally, according to the odd and even before the algorithm / back-end computing a gap between the proposed solving methods, the software model is validated. Three, the design of the main controller for multi emission the RTL code, and the system hardware integration. Firstly, using ModelSim simulation to verify the sub task size 4096:1 scale under the functional correctness, then in the FPGA development board finish Full scale experiments to verify the accuracy of the model and the feasibility of software design to obtain performance information. This improved design of the transmission system has the following characteristics: (1) the task scheduling method is suitable for high density type calculation algorithm, multi emission mode applied in the multi-channel parallel computing method is of high density (2) support. The transmission channel for different kinds of calculation. (3) for computing multiple transmit sub tasks, using dynamic scheduling methods. (4) multi transmitting channel resource sharing system, the number of channels can be extended according to requirement. (5) according to the different transmission channel dynamic optimal allocation of cluster resources. (6) Streamline configuration information length. Dynamic address mapping real / virtual address space. (7) can choose to open the local order optimization function of multi transmitting.

【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】：TP332

【相似文献】