高级检索
    印 杰 江建慧. 冗余多线程结构的重命名寄存器配对共享分配策略[J]. 计算机研究与发展, 2011, 48(3): 516-527.
    引用本文: 印 杰 江建慧. 冗余多线程结构的重命名寄存器配对共享分配策略[J]. 计算机研究与发展, 2011, 48(3): 516-527.
    Yin Jie and Jiang Jianhui. The Matching Share Allocation Strategy of Rename Registers for Redundant Multithreading Architectures[J]. Journal of Computer Research and Development, 2011, 48(3): 516-527.
    Citation: Yin Jie and Jiang Jianhui. The Matching Share Allocation Strategy of Rename Registers for Redundant Multithreading Architectures[J]. Journal of Computer Research and Development, 2011, 48(3): 516-527.

    冗余多线程结构的重命名寄存器配对共享分配策略

    The Matching Share Allocation Strategy of Rename Registers for Redundant Multithreading Architectures

    • 摘要: 同时多线程处理器允许多个线程同时执行,一方面提高了处理器的性能,另一方面也为通过线程冗余执行来容错提供了支持.冗余多线程结构将线程复制成两份,二者独立执行,并比较结果,从而实现检错或者容错.冗余多线程结构主要采用ICOUNT调度策略来解决线程间资源共享问题.然而这种策略有可能造成“饥饿”现象,并降低处理器吞吐率.提出一种重命名寄存器配对共享分配策略,在运行N个独立线程的结构中,将重命名寄存器分成N份,每个主动线程及其相应的冗余线程共享其中的一份,这样就可以比较有效地缓解竞争式共享所带来的负面影响.实验表明,配对共享策略使得处理器的吞吐率和单个线程的性能均有较大幅度的提高.

       

      Abstract: The simultaneous multithreading technique permits multiple issues from different threads simultaneously, which provides nature support for fault-tolerance by executing threads redundantly. Redundant multithreading (RMT) copies each thread into two copies that performances independently, and the corresponding results from the two threads are compared for detecting or tolerating faults. Most RMT structures adopt ICOUNT thread scheduling, which may suffer from long instruction latency. Long latency instructions may stay in critical resources (such as rename register file, instruction queue) for a long time with no contribution to the performance, and the other threads cannot obtain enough resources at the same time. So critical resources cannot be used efficiently, which causes the throughput and the single thread performance degradation. In the paper, a novel rename register allocation method called matching share (MS) is proposed. In N-contexts processor, the rename register file is divided into N copies, one of which is shared by a master thread and its corresponding slave thread, which may avoid the competition of rename register among different threads, and the sharing between master threads and slave threads avoids the drop of resource utilization. The experimental results show that the proposed MS strategy enhances both throughput and performance of single thread.

       

    /

    返回文章
    返回