【作者】 单征；

【导师】 赵荣彩；

【作者基本信息】 解放军信息工程大学 ， 计算机软件与理论， 2007， 博士

【摘要】 由于网络处理器(Network Processor,简称NP)具备高性能、可编程、易扩展和开发周期短等特性,自90年代末诞生伊始就受到了众多商业企业和科研机构的关注。30余个厂商生产或提出了500多种产品和设计方案；众多公司(例如：Cisco、ALCATE、Deceng、华为3Com等)采用网络处理器构建了路由、交换、安全等领域的多种网络设备。近年来,网络处理器技术的发展不尽人意,特别是在利用网络处理器进行应用系统开发时,面临着缺乏能在异构的软硬件体系结构间进行合理的系统设计方案评估,并为网络处理器系统的设计和优化提供技术支撑和参考依据的性能分析方法。本文的主要研究目的是：开展网络处理器系统的性能分析方法研究,为网络处理器系统研发中的方案评估、系统设计、系统改进等阶段提供技术支撑和参考依据。即对众多的器件和系统设计方案进行评估,选择适合需求的性价比好的方案；分析已有的方案和系统的性能缺陷和瓶颈,并根据这些因素改进和提高系统性能；对将要设计的系统进行性能预测,在性能成本方面实现最佳设计和配置。本文的主要贡献与创新总结如下：1、提出了扩展的通用网络处理器结构(Extended General Network Processor Topology,简称EGNPT)和网络处理器软件分层描述法,实现了不同网络处理器软硬件体系结构的统一描述以及网络处理器的软件结构和多级并行层次间的映射,为在不同软硬件结构的网络处理器系统间进行性能分析工作奠定了基础。2、提出了将“可用计算周期”和“访存次数”作为约束条件的网络处理器软硬件结构映射方法,压缩了网络处理器应用设计空间(Design Space)(针对基准测试用例IPv4Forward的实验结果表明该方法的压缩比例高达8947848倍),有效降低了设计空间内搜索的复杂度,形成了合理的性能分析对象集合。3、提出了“基于分层架构的网络处理器系统性能分析方法”,实现了对网络处理器这种多级并行系统的性能评估。实验结果表明该方法的误差小于8%,高于现有各种性能分析方法的精度(误差为10%—17%)。NPF基准测试用例IPv4 Forward的设计方案评估、“基于网络处理器的网络测试平台”的系统设计、“基于网络处理器的网络监控系统”的系统改进等工程实践证明：该方法可以有效的为网络处理器系统的方案评估、系统设计和系统改进等工作提供技术支撑和参考依据。4、在应用功能层面,提出了“线段分析法”,实现了对网络处理器处理单元(Processing Element,简称PE)中多线程程序的性能分析。针对IPv4 Forward的实验结果表明该方法误差仅为0.47%,为进一步采用排队网络进行系统平台层面的性能分析提供了高精度的所需参数。5、在系统平台层面,提出了“基于GI/G/m/∞/FCFS多类型客户开放排队网络”的性能分析模型,利用非乘积形式解排队网络(Non-Product-Form Queueing Networks)描述网络处理器系统,实现了对多类型输入、多服务员服务台的开放网络的性能分析。相对于现有研究成果(仅采用单一客户类型、单服务员服务台和泊松到达过程/负指数服务时间分布的排队网络进行建模),该模型提高了对现实系统的描述贴近度。6、采用分解法对GI/G/m/∞/FCFS多类型客户输入开放排队网络求解,给出了一种新型流通阶段(Flow)输出流到达时间间隔变异系数(Coefficient of Variation)的求解方法。基于该方法设计了“基于排队网络的包随机发送算法”,并应用到河南省杰出人才创新基金项目“基于网络处理器的网络测试平台”中,实现了高速随机测试流量的生成。本文在网络处理器软硬件结构抽象描述及其映射、设计空间压缩、具有多级并行机制的网络处理器系统性能分析方法、基于排队网络的性能分析模型等方向做了较为深入细致的研究,但仍存在有待进一步研究的问题,包括：在不同体系结构的网络处理器平台上验证本文提出的“基于分层架构的网络处理器系统性能分析方法”的普适性；研究网络处理器的指令级并行和程序/数据的相关性对网络处理器系统性能的影响；基于本文提出的性能分析方法研发自动化性能分析工具。更多还原

【Abstract】 From the birth of network processors (NP), they gain lots of attentions from many companies and researchers, because of their high performance, programmable, extendable and fast to markets characteristics. More than 30 companies have brought out more than 500 products or designs. Cisco, ALCATE, Deceng, HuaWei 3Com etc. have built routers, switches and security devices using these processors.But now, the development of network processors technology doesn’t match people’s wish. Especially when building a system with NPs, we have to face the fact that the absence of a performance analysis methodology, which can help us evaluate the various designs among different hardware and software architectures, design and optimize NP-based systems.The main purpose of this dissertation is to solve above problems by the study of performance analysis of NP-based systems. And the dissertation makes the following contributions:1. The dissertation presents the extended general network processors topology (EGNPT) and level based NP software depiction, which solve the problem of how to depict the different NPs’software and hardware architectures in universal form. The dissertation also realizes the mapping of the NPs’software architecture to multilevel parallelism. This work builds the base for the performance analysis of the NP-based applications with different software and hardware architectures.2. The dissertation brings out "compute budget" and "times of memory accessing" restrictions when the software being mapped to hardware. These restrictions reduce the number of mapping set’s (design space) elements and the complexity of the searching working. The constricting test of IPv4 Forward design space shows the original size is 8947848 times of the reduced size. This work makes the propriety objects for NP-based systems performance analyzing.3. The dissertation presents a hierarchical performance analysis methodology for network processor applications, which overcomes the difficulties brought by NP’s multilevel parallelism characteristic. The test results show that the error of this methodology is less than 8% and other performance analysis methods, whose errors are 10%-17%. The practice of the methodology in "IPv4 Forward", "NP-based Network Testing Platform" and "NP-based Network Inspecting System" shows that it can provide technology supports and references for the NP-based system’s performance evaluating, designing and optimizing.4. On the application/function level, the dissertation presents a line based analysis methodology to solve the multithreads program performance evaluation. The test results of IPv4 Forward show that the error of this methodology is 0.47%. This work provides highly accurate parameters for the performance model on system/platform level.5. On the system/platform level, the dissertation models the system using a GI/G/m/∞/FCFS queueing network, which solves the performance analysis of multiclass inputs and multiserver node open queueing network. The model depicts the system more closely to the real world than other models, which use one class customer, Poisson arrival process and Exponential serve time distribution queueing network to model the system.6. The dissertation uses decomposition to solve the GI/G/m/∞/FCFS muliclass open queueing network, and gives a new formula for the coefficient of variation of the interdeparture times. Based on this method, the dissertation designs a queueing network based packet random sending algorithm and implements it in the "NP-based Network Test Platform".The dissertation make efforts in NP software and hardware depictions and mapping relations, NP’s multilevel parallel program performance analysis and queueing networks based performance model. Further researches will be needed, including:a. more testing and evaluating of difference applications on various hardware platforms to validate the universality of the dissertation’s methodology.b. further studies of NP’s instructions parallelism and more analysis of program/data relativity.c. building performance analyzing software tools with the methodology of this dissertation.更多还原