【作者】 徐晓飞；

【导师】 康继昌；

【作者基本信息】 西北工业大学 ， 计算机应用技术， 2004， 博士

【摘要】 本论文的研究课题是航空科学基金项目“面向航空电子综合系统的光纤传输交换技术研究”(02F53031)的组成部分。国外第四代航空电子综合系统的数据通信网络是一个基于光互连的机群系统。开展面向新一代航空电子综合系统的高速通信网络的研究,对我国国防建设具有很重要的现实意义。本文将围绕机群互连网络的通信协议展开研究,主要研究工作包括以下几个方面: 目前常用的机群互连网络,通常采用数据包实现软件通信协议的控制,通信开销较大。本文提出并设计了一种信令控制交换网络TC-NET(Token Control Network), TC-NET通过通信协议中的多层次控制信令,对通信协议进行控制,有效地提高了数据通信效率。 在千兆以太网等商业网络中,通信描述符通常都采用软件的方法实现,每次数据通信过程中,网卡都至少需要两次访问物理内存:第一次从内存中获取通信描述符,然后再次访问内存,取得用户数据,因此通信开销较大。本文采用软硬件协同的设计思想,提出在网卡中设置通信描述符硬件,来支持数据通信。网卡在数据通信过程中,只需访问一次内存,从而提高了通信效率。TC-NET的通信延迟为4微秒,而千兆以太网的通信延迟为19微秒。 作为机群系统快速通信的标准,VIA已经在业界获得了广泛地支持。本文采用软硬件协同的设计思想,提出了支持VIA标准的用户层传输机制TC-VIA(Token Control Virtual Interface Architecture)设计方案,来实现信令控制交换网络TC-NET的用户层通信协议。TC-VIA通过基于Linux底半函数的核心代理机制,一方面在通信的发送端和接收端实现了零拷贝数据通信,另一方面实现了软件VI队列的管理。与采用全软件方法实现的VIA相比,TC-VIA具有更好的通信性能;与采用全硬件方法实现的VIA相比,TC-VIA具有很好的灵活性和可扩展性。 提出在互连网络中采用硬件的方法来实现动态优先级实时通信协议,构成EDF(Earliest Deadline rirst)实时通信网络,满足航空电子综合系统对实时通信的需要。以太网等商业网络通常采用软件方法实现实时调度,支持实时通信,通信延迟通常在毫秒级。本文提出的EDF实时通信网络通过硬件逻辑实现实时调度,可提供小于10微秒的实时通信延迟,相对于以太网等商业网络,EDF实时通信网络通信性能有很大的提高。美国的RaceWay实时通信网络采用硬件的方法实现了静态优先级调度协议,它是作者所查资料中,实更多还原

【Abstract】 The research work of this dissertation is a part of the Aviation Science Foundation Project, named "A Study on the Avionics Oriented Switch Fabric Technology" (02F53031). In the foreign fourth generation avionics systems, the core of data transmitting network is a cluster system based on switch-fabric interconnects. It is of great immediate significance for national defense construction to study high-speed data transmitting network for the next generation avionics in China. This dissertation focuses on the communication protocol in cluster interconnection system. The details of the dissertation is as follows:Presently, the overhead of data communication control mechanism used in cluster system, which is usually realized by data packet, is very high. We have designed and implemented a token control network (TC-NET). The TC-NET utilizes multi-hierarchical control tokens to control the procedure of communication. Consequently, the communication efficiency was increased obviously.In commercial network, such as gigabit Ethernet, the communication descriptor is usually realized by software. Therefore, in each message transfer, the NIC has to access physical memory at least twice: (1) the NIC gets communication descriptor from physical memory, (2) it accesses physical memory to obtain user message. So the communication overhead in commercial network is high. In this research, by taking advantage of Software/Hardware Co-design, we put hardware communication descriptor in NIC to support data communication. With this support in hardware directly, the NIC need to access physical memory only once for each message transfer. Thus the communication efficiency is improved. The communication delay in TC-NET is 4 microseconds and 19 microseconds in gigabit Ethernet.As a fast communication standard of cluster systems, VIA has obtained widely supports in industry. In TC-NET, we designed and implemented the TC-VIA user-level communication mechanism based on VIA standard through an approach of the software/hardware co-design. The kernel agent in TC-VIA is based on Linux bottom half function. By the kernel agent mechanism, the TC-VIA realizes the zero-copy data communication and the management of software VI queue. Comparing to the implementation of VIA based on software purely, the TC-VIA has a better communication performance; comparing to theimplementation of VIA based on pure hardware, the TC-VIA has a good flexibility and extensibility.We have presented and implemented the dynamic priority real-time communication protocol in interconnection network by hardware method and formed EDF real-time communication network. The communication schedule in real-time system based on commercial network is usually realized by software method and the communication delay is usually measured in milliseconds. We implement real-time scheduling based on the EDF algorithm by hardware logic, and the communication delay is less than 10 microseconds. So comparing to commercial network, such as Ethernet network, the communication performance in the EDF real-time communication network has been improved greatly. The American Raceway real-time communication network has the highest performance in literature we have seen. It also realizes the priority schedule by hardware method, but it is based on static priority schedule. Comparing to Raceway network, the EDF network has a better communication flexibility and throughput. Thus, the EDF network can support irregular message flow communication better.In complex real-time communication network, the hard real-time communication task needs to get multiple communication links at the same time in a real-time data transfer. To realize hard real-time communication in complex real-time communication network, the dissertation puts forward MEDF algorithm (Multi-resource Earliest Deadline First) for feasibility-test of message-flow. The message-flow satisfying the feasibility-test can get hard real-time communication servers in EDF real-time communication network. The complexity analysis of MEDF algorithm shows that the complexity of MEDF algorithm is lowest in both time and space.更多还原