面向可重构片上网络的多目标联合优化映射关键技术研究

Mapping optimization is one of the most important research topics of Network-on-Chip (NoC), as different mapping paterns directly affect the NoC and SoC in terms of latency, energy consumption and reliability. Reconfigurable NoC is a promising interconnect structure of Multiprocessor System-on-chip(MPSoC). Based on the communication characteristics of each application, the topology and the routing algorithm of reconfigurable NoC can be changed accordingly to satisfy the specific requirements. The unique characteristics of reconfigurable NoC bring along new challenges and requirements for NoC mapping optimization. This proposal focuses on the flexibility (i.e. being applicable to various topologies and routing algorithms), efficiency (i.e. for re-mapping after dynamic reconfiguration), and accuracy (i.e. cost model involving multiple metrics including latency, energy and reliability). We propose a cost model considering the statistics of all possible cases of links being faulty with independent possibilities to enhance the accuracy. The optimal solution searching method is explored by optimizing the branch and bound algorithm with the priority allocation and a compensation factor. The research results of the project can be directly applied to NoC-based reconfigurable computing architecture as well as MPSoC, acting as a strong support for developing indigenous high performance MPSoC.

映射优化是片上网络研究的一项关键技术，不同的映射方式将影响片上网络以及整个片上系统的延时、能耗、可靠性等一系列指标。为了满足片上多核系统针对不同应用的性能要求，可重构片上网络能够根据不同的通信特征进行硅后定制化配置。这一特征对片上网络映射优化提出了新的挑战和需求。本课题着眼于综合解决可重构片上网络所需的适合于不同拓扑结构和路由算法的灵活性、满足动态改变配置之后重新映射的高效性、以及考虑延时、能耗和可靠性多目标模型的准确性这三个问题。通过引入优先级和补偿因子的分支界定方法保证准确性的前提下大幅减少求解最优模型的计算复杂度，通过概率计算所有互连线发生独立错误概率的情况提高多目标建模的准确性。本课题的研究将可直接应用到基于可重构片上网络互连的多处理器片上系统以及可重构计算处理器，将为我国未来自主研发高性能片上多核系统提供战略技术支撑。

映射优化是片上网络研究的一项关键技术，不同的映射方式将影响片上网络以及整个片上系统的延时、能耗、可靠性等关键指标。为了满足片上多核系统针对不同应用的性能要求，可重构片上网络能够根据不同的通信特征进行硅后定制化配置。这一特征对片上网络映射优化提出了新的挑战和需求。针对如何在片上网络上实现算法映射优化以同时满足动态可重构高效性、准确性和灵活性需求的问题，提出了同时定量分析可靠性、通信功耗和延时的可靠性效率模型，通过考虑所有互连线发生错误的所有可能情况来提高模型的准确性，以及基于优先级分配和补偿因子的改进分支界定算法。通过优先级分配的方法大大提高了搜索效率，通过引入补偿因子从而实现在提高寻找效率的同时确保准确率不受影响。该方法求解最优映射的速度比传统的分支界定方法和模拟退火方法，分别提升了3倍和720倍,且找到的最优映射法平均可靠性更高、通信能耗更小、性能更好。本项目的研究成果将可应用到基于可重构片上网络互连的多处理器片上系统以及高能效可重构计算处理器，为我国自主研发高性能片上多核系统提供战略技术支撑。