基于x86的高速报文处理技术研究与实现

发布时间：2018-10-17 08:20

【摘要】：随着“互联网+”渗透到人们生活各方面,其诸多问题日益突出。一方面期望网络速率更高,另一方面需要处理海量网络数据以解决网络优化等问题。高速报文处理软件的性能出现瓶颈,尤其在多核平台和Linux操作系统结合时。已经有越来越多的热点研究成果出现,包括Linux自身对报文处理机制改进和迭代。PF_RING、DPDK、NET-MAP是同时期诞生的高速报文处理框架,它们在技术和性能特点上互有异同。所以用户在框架选择和性能判断上也存在困难。已有文章已经对PF RING、DPDK、NET-MAP进行了多个维度的性能对比。但这方面研究还不能完全满足实际需要,且覆盖面也窄。问题包括未考虑纯报文接收场景、NUMA平台上的性能特性对比不足、多核心扩展问题研究不足等。本文首先论述并分析了现有Linux报文处理机制和关键优化技术,重点分析了 PF_RING和DPDK的优化技术异同点。然后对PF_RING和DPDK对比研究进行展开。首先简要说明当前现有研究对其研究的不足,介绍了 Run-to-completion和Pipeline模型(样式)分类,然后提出对不同框架的对比研究需要从为两模型入手并选择其中的几种精简模型做作析。本文提出一种报文处理研究分析的理想化模型——流阻分析理想模型。结合流阻模型对PF_RING和DPDK的Run-to-completion和Pipeline的精简模型进行整体性能、不同负载性能、多核性能及可扩展性进行分析和评估。再将性能问题的研究扩展到NUMA平台上,根据NUMA平台的特殊性,将PF_RING和DPDK的Run-to-completion和Pipeline的精简模型进行针对NUMA平台的分析。由此研究得到了不同框架在NUMA平台的性能特性,特别是其跨NUMA节点可扩展性和线程在NUMA节点间的分布样式对性能的影响。最后基于现有的报文处理优化技术设计了一套离线报文高速回放系统,并对回放程序中的多个参数的性能作了性能分析和优化。最后一章对全文总结并说明本文的不足和未来研究的展望。
[Abstract]:With the penetration of the Internet into all aspects of people's lives, its many problems become increasingly prominent. On the one hand, the network speed is expected to be higher, on the other hand, we need to deal with massive network data to solve network optimization problems. The performance of high-speed packet processing software appears bottleneck, especially when multi-core platform and Linux operating system are combined. There have been more and more hot research achievements, including the improvement and iteration of packet processing mechanism by Linux itself. PF_RING,DPDK,NET-MAP is a high-speed packet processing framework born at the same time, and they have different technical and performance characteristics. Therefore, users also have difficulties in frame selection and performance judgment. Previous articles have compared the performance of PF RING,DPDK,NET-MAP across multiple dimensions. However, the research in this area can not fully meet the actual needs, and the coverage is also narrow. The problems include not considering the pure message receiving scenario, the lack of performance comparison on NUMA platform, and the lack of research on multi-core extension. This paper first discusses and analyzes the existing Linux packet processing mechanism and key optimization techniques, and focuses on the similarities and differences between PF_RING and DPDK optimization techniques. Then the comparative study of PF_RING and DPDK was carried out. Firstly, the deficiency of the current research is briefly explained, and the classification of Run-to-completion and Pipeline models (styles) is introduced. Then, it is proposed that the comparative study of different frameworks should start with two models and select several reduced models for analysis. In this paper, an idealized model for packet processing and analysis is presented, which is an ideal model for flow resistance analysis. Combined with the flow resistance model, the overall performance of the Run-to-completion and Pipeline models of PF_RING and DPDK is analyzed and evaluated. The performance of different load, multi-core and extensibility are analyzed and evaluated. Then the research of performance problem is extended to NUMA platform. According to the particularity of NUMA platform, the reduced model of Run-to-completion and Pipeline of PF_RING and DPDK is analyzed for NUMA platform. The performance characteristics of different frameworks on NUMA platform are obtained, especially the influence of extensibility across NUMA nodes and the distribution style of threads among NUMA nodes on the performance. Finally, an off-line high-speed playback system is designed based on the existing packet processing optimization techniques, and the performance of multiple parameters in the playback program is analyzed and optimized. The last chapter summarizes the full text and explains the shortcomings of this paper and the prospects for future research.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP393.0;TP316.81

【相似文献】