面向多核网络分处理系统的线程亲和缓冲区管理机制（.docVIP

面向多核网络分组处理系统的线程亲和缓冲区管理机制? ?收稿日期：2015-05-27 基金项目：国家自然科学基金资助项目（612024823； 作者简介：杨惠（1987－），女，安徽萧县，助理研究员，博士，E-mail：huihu126.com； 杨惠1，陈一骄1，李韬1，李世星1，戴幻尧2 （1.国防科学技术大学 计算机学院，湖南 长沙 410073；2.中国洛阳电子装备试验中心，河南 洛阳 471003 ） 摘要：基于通用多核架构的网络分组处理系统性能受到诸如分组IO开销高、多核共享内存及进程调度竞争大、TLB表项失效率高等问题的困扰。为此提出一种基于通用多核网络分组处理系统，面向高速分组转发应用的线程亲和缓冲区硬件管理机制，并在网络专用协处理引擎上实现。该机制采用无中断的线程亲和调度策略，将包含控制信息与缓冲区地址信息的描述符和分组数据，按照分组处理的线程号，链式的对应加载在多个地址连续的共享缓冲区中。基于通用多核和FPGA平台进行报文转发测试，实验结果表明，采用线程亲和缓冲区管理机制，能使平均报文转发处理性能提升12.4%，有效地降低IO开销和TLB表频繁缺失的问题。 关键词：分组IO；线程亲和；缓冲区管理；分组转发；多核 中图分类号：TP393 文献标志码：A　　 文章编号： Thread Affinity for Buffer Management Mechanism based on Multi-core Network Processing System YANG Hui1, CHEN Yijiao, LI Tao, LI Shixing, DAI Huanyao2 (1. College of Computer, National University of Defense Technology, Changsha 410073, China；2.Luoyang Electronic Equipment Test Center，Henan, Luoyang 471003, China) Abstract: The packet processing performance of the general multi-core is plagued by many factors, including high packet I/O cost, multi-core share memory competition, TLB entries failure rate, etc. A thread affinity for buffer management mechanism TABM based on multi-core network packet processing system for high-speed packet forwarding application is proposed and completed on network dedicated co-processing engine NPE. TABM mechanism adopts thread affinity scheduling strategy with no interrupt, it sends each packet data and descriptor which contains control and buffer address information to several successive shared buffers according to the corresponding thread ID, and organizes the packets and descriptors which processed in the same thread in the form of a chain. This paper tests the packet forwarding performance based on general multi-core and FPGA platform. The experimental data shows that TABM can obtain the packet forwarding performance speedup about 1.12x over the baseline architecture. It can effectively reduce the IO overhead. Keywords: packet IO; thre