基于通用多核平台的高速网络处理系统关键技术研究

发布时间：2018-05-07 19:53

本文选题：通用多核处理器 + 高性能网络处理程序　；参考：《中国科学技术大学》2014年博士论文

【摘要】：随着万兆链路进入实际部署以及数据包处理功能迅速向协议栈高层转移,网络系统的处理能力正面临越来越严峻的挑战。传统上,高速网络设备主要依靠专用硬件(如ASIC、网络处理器)达到高性能。但是硬件方案可编程性差,只能承担简单的包处理功能。随着网络应用日趋复杂,网络系统的处理任务正从简单的包转发、包过滤演变到深度数据包检查、网络异常检测、基于内容的交换、数据加密等复杂功能。这种深度的包处理要求网络系统提供更多的计算能力和编程灵活性。自双核处理器出现后,通用多核处理器的发展令人瞩目。目前通用多核处理器的核数已经接近、甚至超过了网络处理器中计算单元的数量,其高层协议处理能力更是专用硬件无法比拟的。基于通用多核平台开发网络设备已成为目前学术界和工业界感兴趣的研究课题。然而通用多核平台并不是为网络处理而设计,将它应用于高速网络系统面临诸多难题,比如缺乏高效的核间通信机制,通用操作系统的网络子系统不够高效,网络程序的并行化方法没有先例可借鉴,等等。本论文系统研究基于通用多核平台构建高速网络系统的若干关键技术,并在Intel通用多核平台和Linux操作系统上实现网络系统原型。本论文从构建一个具体的网络流量监视系统入手,研究并实现流水线并行模型、高效的核间通信机制、基于无锁设计的全局数据结构、串行代码到并行代码的转换等关键技术。本论文的主要贡献及创新点如下： (1)结合流水线并行和任务级并行模型,在Intel通用多核平台上实现了一个高性能的流量监视系统原型。该系统原型可达到2Gbps应用层处理能力。 (2)针对流水线并行模型中最关键的流水线级间通信问题,设计并实现了一个简单、高效且实用的单生产者／单消费者并发无锁队列B-Queue。B-Queue性能与已有同类算法相同,且在真实环境中性能稳定。 (3)针对全局数据结构共享难有有效解决方案的问题,提出面向特定应用设计定制化无锁数据结构的解决思路和方法,并运用到一个具体的TCP SYN洪泛攻击检测应用程序中,为将无锁设计思想运用到具体的程序设计实践中进行了有益的探索。本论文工作为利用通用多核服务器建立高速网络系统的研究与实践积累了宝贵的经验。
[Abstract]:With the actual deployment of the multi-megabit link and the rapid transfer of the packet processing function to the high layer of the protocol stack, the processing capability of the network system is facing more and more severe challenges. Traditionally, high-speed network devices mainly rely on dedicated hardware (such as ASIC, network processor) to achieve high performance. However, the hardware program is not programmable, and can only undertake simple packet processing functions. With the increasing complexity of network applications, the processing tasks of network systems are changing from simple packet forwarding, packet filtering to deep packet checking, network anomaly detection, content-based exchange, data encryption and other complex functions. This depth of packet processing requires more computing power and programming flexibility for network systems. Since the emergence of dual-core processors, the development of universal multi-core processors has attracted much attention. At present, the number of cores of general multi-core processors is close to, even more than the number of computing units in network processors, and its high-level protocol processing capability is incomparable to that of dedicated hardware. Developing network equipment based on general multi-core platform has become an interesting research topic in academia and industry. However, the general multi-core platform is not designed for network processing. The application of the platform to high-speed network systems faces many difficulties, such as the lack of efficient inter-core communication mechanism, the lack of efficient network subsystem of the general operating system. There is no precedent for parallelization of network programs, and so on. In this paper, some key technologies of constructing high speed network system based on general multi-core platform are studied, and the prototype of network system is implemented on Intel platform and Linux operating system. Starting with the construction of a specific network traffic monitoring system, this paper studies and implements pipeline parallel model, efficient inter-core communication mechanism, global data structure based on lock-free design, and the conversion from serial code to parallel code. The main contributions and innovations of this thesis are as follows: 1) combining pipelined parallelism and task-level parallelism, a high-performance traffic monitoring system prototype is implemented on Intel universal multi-core platform. The prototype of the system can achieve 2Gbps application layer processing capability. In order to solve the most critical problem of pipeline level communication in pipeline parallel model, a simple, efficient and practical single producer / consumer concurrent lock-free queue B-Queue.B-Queue is designed and implemented, which has the same performance as existing algorithms. And the performance is stable in the real environment. In order to solve the problem that it is difficult to share global data structure effectively, this paper puts forward the solution of customizing lock free data structure for specific applications, and applies it to a concrete TCP SYN flood attack detection application program. In order to apply the lock-free design thought to the concrete programming practice, this paper makes a beneficial exploration. This thesis has accumulated valuable experience for the research and practice of establishing high-speed network system using universal multi-core server.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP393.05

【共引文献】