基于图形处理单元的高性能网络包处理技术研究

Packet processing is one of the key network tecnologies, and promoting its performance has vast importance to realize high-speed data interconnection and interactive. Based on research basis of packet processing, common data structures and algorithms, this project, by utlizing Grphic Processing Units (GPUs)- a new accelerating platform, focuses on the research of high-performance routing lookup, packet classification and filtering. By deeply mining the massive parallelsim and vast computing power of GPU, by combining the charactersitics of thread scheduling and memory accessing, this project makes a great of efforts on optimizing the design of mechansims, policies, structures and algorithms, to improve the comprehensive performance of packet processing. This project explores a flexible multi-dimensional mechansim of batch processing, by incorporating GPU's architecture features and the multi-stream pipeline. Moreovere, effective structures and algorithms, targeted to the GPU platform, are studied for new network applications and architecture, such as the virtual router platform, Named Data Networking and so on. Hopefully, the research results could be used in next-generation high-performance network routing/switching and security devices. Specifically, for some complex processing processes of packets, this project proposes a "compressing first and then accelerating" method, which also has great reference value for accelerating other similar problems using GPU. Hence, this project has bright application prospect and important academic value.

网络数据包处理是互联网的关键技术之一，有效提升网络包处理的性能对实现高速数据互联及交互具有重要意义。本课题基于包处理机制以及常用数据结构和算法的前期研究，面向新型加速平台- - 图形处理单元（Graphic Processing Unit, GPU），围绕高性能路由查找及数据包分类过滤技术，深入挖掘GPU的并行计算能力，并结合其调度、访存等方面的特性来优化机制、策略以及结构和算法的设计，以实现综合性能的提升。本课题不仅结合GPU的架构特性和多流技术探索灵活的多维度批处理触发机制，且面向虚拟化路由器平台、命名数据网络等新型应用场景和网络架构研究适合GPU平台的包处理结构和算法。所取得的研究成果将有望应用到下一代高性能网络路由/交换及安全设备中。尤其是本课题针对复杂包处理问题提出的"先压缩再加速"思路，对类似问题的GPU加速研究都具有重要的参考意义。因此，本课题的研究具有良好的应用前景和学术价值。

网络数据包处理是互联网的关键技术之一，有效提升网络包处理的性能对实现高速数据互联及交互具有重要意义。本课题基于包处理机制以及常用数据结构和算法的前期 研究，面向新型加速平台——图形处理单元(Graphic Processing Unit, GPU)，围绕高性能路由查找及数据包分类过滤技术，深入挖掘 GPU 的并行计算能力，并结合其调度、访存等方面的特性来优化机制、策略以及结构和算法的设计，以实现综合性能的提升。本课题不仅面向IPv6、虚拟化路由器平 台、命名数据网络等新型应用场景和网络架构研究适合 GPU 平台的包处理结构和算法，还从整体系统层面探索CPU/GPU混合的协同包处理架构。项目组在IPv6地址查找、命名数据网络名字查找、虚拟化路由器多表查找、5元组包分类以及多元组的OpenFlow流分类等方面设计了一系列新型数据结构和算法，以配合GPU多核加速。同时，项目组还在数据包IO、内存管理、任务调度等方面设计了一系列高效机制实现CPU/GPU的协同处理及性能和能效优化。所取得的研究成果将有望应用到下一代高性能网络路由/交换及安全设备中。