支持混合计算模式的异构多核系统若干关键技术的研究

发布时间：2018-12-24 21:04

【摘要】：以图形图像、科学计算、大数据分析等应用为代表的高密度计算应用领域,具有数据吞吐量大、实时性要求高的特点,对微处理器的处理能力、数据吞吐率以及处理并行度等均提出了更高的要求。多核技术将处理器的发展方向从单一的纵向发展扩充了横向发展,大大降低了达到同种计算能力时处理器的设计难度,也缓解了功耗等因素对处理器设计的制约,在近十余年中迅速发展为处理器设计的主流方向;可重构计算兼顾了定制计算的高效性与通用计算的灵活性,是解决高密度计算领域计算需求的一种优秀的计算体系结构;动态调度技术,作为当代处理器设计中的一种重要技术,能够消除任务间的数据伪相关,提高处理器的发射效率,从而显著提升处理器的工作能力。在计算中,有两种典型的数据计算结构:存储计算方式和流计算方式。存储计算方式便于实现数据的重用,却需要占用大量时间来完成数据搬运;流计算方式能够隐藏数据搬运时间,大大提高计算吞吐率,却会对数据存储和交换带来较高的带宽压力。基于以上背景,论文对支持混合计算模式的异构多核系统的关键技术进行了讨论与研究,主要工作如下:首先,论文对原有基于多重片上网络的异构多核计算系统架构进行了介绍与分析,用两级计算架构思想对目标系统进行了改造,并规范化了多核计算系统的工作机制,设计了顶层任务指令的指令集;其次,论文设计了一种用于多核计算系统任务调度的主控制单元,引入了动态调度和寄存器重命名技术,实现了任务指令级的乱序多发射,并进一步探索了线程级并行的实现方案,提高了目标系统的任务发射效率;再次,论文设计了一种粗粒度可重构计算单元,采用可重构技术设计,支持多种计算模式,并对高密度计算领域中常见的算法进行了定向优化,在有限资源条件下达到了较高的计算能力;最后,本文对设计和改进后的单元进行了测试和性能评估,验证了设计的正确性,并讨论了算法映射中需要注意的问题,提出了系统进一步优化的方向。
[Abstract]:The applications of high density computing, such as graphics, image, scientific calculation, big data analysis and so on, have the characteristics of large data throughput and high real-time requirements, and have the ability to process microprocessors. Data throughput and processing parallelism are higher requirements. Multi-core technology expands the development direction of processor from single vertical development to horizontal development, greatly reduces the difficulty of processor design when the same computing power is achieved, and also alleviates the restriction of power consumption and other factors on processor design. In the past ten years, it has developed rapidly into the mainstream of processor design. Reconfigurable computing takes into account the high efficiency of customized computing and the flexibility of general computing. It is an excellent computing architecture to solve the computing requirements in the field of high-density computing. As an important technology in the modern processor design, dynamic scheduling technology can eliminate the data pseudo-correlation between tasks, improve the efficiency of the processor, and improve the working ability of the processor. There are two typical data computing structures: storage computing and stream computing. The method of storage and calculation is convenient to realize the reuse of data, but it takes a lot of time to complete the data handling. Stream computing can hide the data handling time, greatly improve the computational throughput, but will bring higher bandwidth pressure to data storage and exchange. Based on the above background, this paper discusses and studies the key technologies of heterogeneous multi-core systems supporting hybrid computing mode. The main work is as follows: first, This paper introduces and analyzes the original architecture of heterogeneous multi-core computing system based on multi-chip network, reconstructs the target system with the idea of two-level computing architecture, and standardizes the working mechanism of multi-core computing system. The instruction set of top-level task instruction is designed. Secondly, a main control unit is designed for task scheduling in multi-core computing systems. Dynamic scheduling and register renaming techniques are introduced to realize the task instruction level disordered multi-transmission. Furthermore, the implementation scheme of thread-level parallelism is explored, and the task launching efficiency of the target system is improved. Thirdly, a coarse-grained reconfigurable computing unit is designed, which uses reconfigurable technology to support a variety of computing modes, and optimizes the common algorithms in the field of high-density computing. Under the condition of limited resources, high computing power is achieved. Finally, this paper tests and evaluates the design and performance of the improved unit, validates the correctness of the design, discusses the problems needing attention in the algorithm mapping, and puts forward the direction of further optimization of the system.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP332

【参考文献】