面向领域应用的可重构系统关键技术研究

发布时间：2018-06-14 15:32

本文选题：领域应用 + 可重构计算　；参考：《解放军信息工程大学》2015年博士论文

【摘要】：当前科学计算领域的处理部件大致可分为通用计算器件和专用计算器件。通用计算器件具有通用性强、但效能不高的特点,专用计算器件则具有灵活性差但效能高的特点。可重构计算器件的兴起成为这两种计算模式的缓冲带,不仅能够通过重构成为完成某项功能的高效专用器件,而且可以通过动态重构灵活地实现不同功能的切换。而目前通用的可重构计算,为了满足通用性、灵活性的要求,其重构粒度相对较细,这样就使得对其管理、调度、重构等代价相对较大。因此,在丰富和发展了可重构理论与技术的基础上,PRCA(Proactive Reconfigurable Computing Architecture)体系结构理念,其引入构件概念对重构粒子进行扩展。因此,从逻辑门、片内逻辑块、IP核、部件直到子系统都是多维重构的“可重构粒子”。在PRCA系统中,基于对“应用的属性、服务质量、负载变化、系统构件资源”的认知,来识别应用的需求、应用的变化,以及系统中可以利用的处理资源。依据尽可能高效、或可信安全的原则,在应用需求与计算资源间动态决策出效能最优的匹配,构建出适合于应用需求的处理结构。并且该结构随着应用的变化,如：计算进展阶段、处理负荷等的变化,而进行结构的主动变更。这样,就实现了结构适应应用的理念。PRCA体系结构具有很多优点,当然也存在相当多的挑战。基于此背景,本文主要解决其在实现过程中的几个关键问题,本文的内容也围绕着这几个问题展开,主要工作及创新点包括：1.研究了基于领域应用的特征分析方法,提出了多维度应用特征模型。对于领域内有限应用的集合,存在着最适宜的处理结构,在满足领域内应用的基础上达到效能最优。而特征是领域应用本质属性的反映,因此基于对领域应用特征的分析,可以建立最合适的计算模型,为使用和构建最合适的处理部件提供必要支撑。特征模型的方法已广泛应用于多个领域,如制造业、自然语言分析、软件设计等。本文把特征模型的方法引入到结构设计当中,建立了以特征向量、特征矩阵模型的应用特征表达形式,给出了特征相关概念的定义,研究了特征的运算方法和相关性质,并建立了从特征树建立、特征量化、特征空间建立到特征规约聚类的领域应用算法建模过程。2.研究了基于算核的应用算法实现方法,提出了算核利用率均等(EUCI)的算核求解算法。基本功能单元设计的好坏,很大程度上决定了体系结构的计算效率。因此,本文基于算法特征,把计算粒化分解为一个个相对独立的子过程,使领域应用划分为具有共性的算核,为提高硬件资源计算的复用程度,使其具有高性能和灵活性,提供了理论基础。把效能控制转化为资源利用率问题,进而表达为算核利用率问题,提出了算核利用率均等迭代算法来求解算核,达到了高效能的目的。3.研究了可重构计算模型,提出了基于元结构的计算模型。计算模型是算法设计者与体系结构之间的一个桥梁,是算法设计和分析的基础。原有的时间、空间二维计算模型受制于主频、功率等因素,即能耗因素。因此引入功耗,作为独立于空间和时间的一维约束,使结构受解算目标的应用驱动,兼顾高效计算和灵活性。并且,为了屏蔽异构处理资源的差异,将硬件结构划分为应用结构和元结构,并论述了元结构的组合和转化形式。为了实现效能计算,在时、空计算的基础上,提出了基于元结构的时空能计算模型,分析了模型的时间、空间、能耗不同维度特性。建立了异构资源的统一结构视图。4.研究了从应用到计算结构的映射方法,提出了基于模糊化效用函数的二人非零和博弈算法。对于具体计算的子任务要分配到由不同类型的处理器构成的异构计算系统中,有多种对应方法。不同的映射结果性能不同,甚至差距很大。本文从模糊化代价函数入手,以既竞争又合作的复杂关系为处理对象,采用博弈论方法,实现算法和结构的多对多映射,提供了一个能平衡任务和资源均衡收益的任务到计算结构的映射算法。实验表明,该方法针对具体应用、环境、运行状况等任务和系统资源参数,达到了平衡应用和资源的各种效益,实现了应用和系统的均衡优化目标。5.从不同测度空间研究了体系结构的评价方法,提出了多维联合评价体系。为了评价多维的计算模型,建立了多维的评价体系,研究了不同维度间评价指标的相互关系。并以多维评价体系为目标,构建了面向领域应用可重构计算的验证系统和验证平台,综合分析了测试结果,论证了面向应用特征的计算方式的正确性和高效能。
[Abstract]:At present, the processing parts in the field of scientific computing can be divided into general calculator parts and special computing devices. The general calculator has the characteristics of strong generality, but its efficiency is not high. The special calculator has the characteristics of poor flexibility but high efficiency. The rise of reconfigurable computing devices is the buffer zone of these two computing modes. By refactoring, it is an efficient and special device to complete a certain function, and can be flexible to switch different functions through dynamic reconfiguration. At present, the reconfigurable computing is relatively fine in order to meet the generality and flexibility, so that the cost of management, scheduling and reconfiguration is relatively large. On the basis of enriching and developing Reconfigurable Theory and technology, the concept of PRCA (Proactive Reconfigurable Computing Architecture) architecture is introduced into the component concept to expand the reconfigurable particles. Therefore, from logical gates, intra chip logic blocks, IP cores, components to subsystems are "reconfigurable particles" of multidimensional reconfiguration. In PRCA systems, the system is a "reconfigurable particle". Based on the cognition of "application attributes, quality of service, load change, system component resources", the application needs, application changes, and the processing resources that can be used in the system are identified. Based on the principle of as high as possible as possible, or trusted security, the optimal matching between application requirements and computing resources is constructed. The structure is suitable for application requirements. And with the change of application, such as the change of computing progress, the change of load and so on, the structure is changed actively. So, the idea.PRCA architecture that realizes the application of structure has many advantages, and of course there are quite a lot of challenges. Based on this background, this paper mainly In order to solve several key problems in the process of implementation, the content of this paper also revolves around these problems. The main work and innovation points include: 1. research the feature analysis method based on the domain application, and put forward the multi-dimensional application feature model. Based on the application of the domain, the performance is optimal. The feature is the reflection of the essential properties of the domain application. Therefore, based on the analysis of the domain application features, the most suitable computing model can be established to provide the necessary support for the use and construction of the most suitable processing components. The method of feature model has been widely used in many fields, such as manufacturing, In this paper, the method of feature model is introduced into the structure design, the feature vector and feature matrix model are established, the definition of feature correlation concept is given, the operation method and correlation quality are studied, and the feature tree is established, the feature is quantified and the feature is set up. The application algorithm modeling process of space based on feature protocol clustering.2. studies the implementation method based on kernel based application algorithm, and proposes a kernel solving algorithm for kernel utilization ratio (EUCI). The design of basic functional units determines the computational efficiency of the architecture to a large extent. Therefore, this paper is based on the algorithm features. The calculation of granular decomposition into a relatively independent subprocess makes the domain application divided into a common kernel. In order to improve the reuse degree of the hardware resource calculation and make it have high performance and flexibility, it provides the theoretical basis. The efficiency control is transformed into the resource utilization rate problem, and then it is expressed as the kernel utilization rate problem, and the calculation kernel is put forward. The utilization ratio iterative algorithm is used to solve the kernel. The reconfigurable computing model is studied by.3.. The calculation model based on the meta structure is proposed. The calculation model is a bridge between the algorithm designer and the architecture. It is the basis of the algorithm design and analysis. The original time, the spatial two-dimensional calculation model is subject to the master. Frequency, power and other factors are the factors of energy consumption. Therefore, power consumption is introduced as one dimension constraint independent of space and time, which makes the structure be driven by the application of the target and takes account of the high efficiency calculation and flexibility. In order to shield the difference of heterogeneous processing resources, the hardware structure is divided into application structure and meta structure, and the combination of the meta structure is discussed. In order to realize efficiency calculation, on the basis of time and space computing, a spatio-temporal energy calculation model based on meta structure is proposed, and the time, space and energy consumption of the model are analyzed. A unified structure view.4. for heterogeneous resources is set up to study the mapping method from application to computing structure, and the fuzzy efficiency is proposed. The two person non zero sum game algorithm of function is used. There are a variety of corresponding methods for the subtasks of specific computation to be assigned to heterogeneous computing systems composed of different types of processors. The performance of different mapping results is different, and the difference is very large. This paper starts with the fuzzy cost function and deals with the complex relationship of both competition and cooperation. By using the method of game theory, the multi to multi mapping of algorithms and structures is realized, and a mapping algorithm which can balance task and resource equilibrium income to computing structure is provided. The experiment shows that the method has achieved various benefits in balancing application and resources for specific applications, environment, operation status and other tasks and system resource parameters. In order to evaluate the multidimensional model, a multi-dimensional evaluation system is set up, and the interrelationship between the evaluation indexes of different dimensions is studied, and the multi-dimensional evaluation system is set up as the goal, and the.5. is constructed. A verification system and a verification platform for domain application reconfigurable computing are presented. The test results are analyzed synthetically, and the correctness and efficiency of the application oriented computing method are demonstrated.
【学位授予单位】：解放军信息工程大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP18

【相似文献】