基于FPGA的高性能计算架构硬件任务与资源模型研究

发布时间：2018-02-26 03:21

本文关键词： FPGA计算加速最大空闲矩形任务调度硬件任务情境硬件资源情境　出处：《上海大学》2012年博士论文　论文类型：学位论文

【摘要】：高性能计算是一个国家的综合国力的体现，是支撑国家实力持续发展的关键技术之一。近年来，高性能计算机体系结构技术研究发生了改变，异构体系结构已成为未来高性能计算机发展的主要趋势。基于FPGA的可重构计算作为一种新的体系结构，让系统拥有了硬件的高性能，又具备了软件的灵活性。通过采用主/协处理器技术，将计算的任务交由计算加速部件以硬件任务完成，而任务管理等，则交由通用处理器来完成，达到一个优化的计算效果。本文主要对基于FPGA计算加速的异构高性能计算架构上的任务与资源管理算法与计算模型方面的研究。在研究与分析当前高性能计算体系结构的发展趋势的基础上，以异构高性能计算平台为研究目标，结合FPGA计算加速，通过对多体问题（N-body）求解的FMM算法在FPGA计算加速的加速效果，通过分析FPGA加速上的计算性能效果，提出了多级加速优化方案与对应的计算架构。资源管理是任务调度研究的基础，通过研究查找空闲矩形空间的算法来遍历这些最大的空闲空间矩形MFR全集，本文分别以基于状态矩阵模型与运行任务边线模型来研究MFR全集查找与管理算法。为有效查找与管理MFR全集，在基于资源状态矩阵模型中提出了基于双向倒形塔的MFR全集扫描求解算法，并在此基础了又给出扫描优化算法与M值标示优化算法。在基于运行任务边线模型上，提出了基于上右边线交点CPTR的全集MFR查找算法，并给出了在线调度时的基于FPGA局部影响空间上的MFR全集更新算法。高性能计算平台多是属于商业应用计算平台，要为众多的高性能计算用户提供计算服务，针对高性能计算平台的多级任务调度模型，提出了基于本地资源FPGA上的时间与空间情境CBTA的多情境状态的硬件任务放置与调度算法体系，并根据设置的不同的任务情境与资源情境状态，提出了多种不同的适应于任务与资源情境状态下的任务调度与放置算法。采用让每个计算资源节点根据自己的资源情境状态变化，而主动去选择对应自己情境的任务的自适应任务调度策略，并给出了CBTA调度算法的并行优化策略。最后通过实验来说明了算法在对用户响应时间、负载均衡以及任务拒绝率上的优势。
[Abstract]:High performance computing is the embodiment of a country's comprehensive national strength and one of the key technologies to support the sustainable development of national strength. In recent years, the research of high-performance computer architecture technology has changed. Heterogeneous architecture has become the main trend of the development of high-performance computers in the future. As a new architecture, reconfigurable computing based on FPGA enables the system to have the high performance of hardware. By using the master / coprocessor technology, the computing task is assigned to the computing acceleration unit to complete the hardware task, and the task management is handed over to the general purpose processor to achieve an optimized computing effect. In this paper, the task and resource management algorithms and computing models of heterogeneous high-performance computing architecture based on FPGA computing acceleration are studied, based on the research and analysis of the development trend of current high-performance computing architecture. Taking heterogeneous high performance computing platform as the research goal, combining with the acceleration of FPGA computation, the acceleration effect of FMM algorithm for solving multibody problem (N-body) in FPGA computation is analyzed, and the computational performance effect on FPGA acceleration is analyzed. A multilevel acceleration optimization scheme and its corresponding computing framework are proposed. Resource management is the foundation of task scheduling. By studying the algorithm of finding free rectangular space, we can traverse these maximal free space rectangular MFR complete sets. In this paper, based on the state matrix model and the running task edge-line model, we study the MFR complete set lookup and management algorithm, in order to find and manage the MFR complete set effectively. Based on the resource state matrix model, the MFR full set scanning algorithm based on the bidirectional inverted tower is proposed, and the scan optimization algorithm and the M value marking optimization algorithm are also given. On the basis of the run-time task boundary line model, the scanning optimization algorithm and the M value marking optimization algorithm are presented. In this paper, a full set MFR lookup algorithm based on the intersection point CPTR of the upper and right line is proposed, and the MFR complete set updating algorithm based on the local influence space of FPGA in the online scheduling is also presented. The high performance computing platform belongs to the commercial application computing platform. It is necessary to provide computing services for many high performance computing users, aiming at the multilevel task scheduling model of the high performance computing platform. This paper proposes a hardware task placement and scheduling algorithm system based on time and space situation CBTA on local resource FPGA, and according to the different task situation and resource situation state, the hardware task placement and scheduling algorithm system based on local resource FPGA is proposed. In this paper, a variety of task scheduling and placement algorithms are proposed, which are suitable for task and resource situation, and each computing resource node is asked to change according to its own resource situation. Meanwhile, the adaptive task scheduling strategy of the task corresponding to their own situation is chosen, and the parallel optimization strategy of the CBTA scheduling algorithm is given. Finally, the response time of the algorithm to the user is illustrated by experiments. The advantages of load balancing and task rejection rates.
【学位授予单位】：上海大学
【学位级别】：博士
【学位授予年份】：2012
【分类号】：TP38;TN791

【参考文献】