高级检索
    田杭沛 高德远 樊晓桠 朱怡安. 面向实时流处理的多核多线程处理器访存队列[J]. 计算机研究与发展, 2009, 46(10): 1634-1641.
    引用本文: 田杭沛 高德远 樊晓桠 朱怡安. 面向实时流处理的多核多线程处理器访存队列[J]. 计算机研究与发展, 2009, 46(10): 1634-1641.
    Tian Hangpei, Gao Deyuan, Fan Xiaoya, and Zhu Yian. Memory Request Queue of Multi-Core Multi-Threading Processor for Real-Time Stream Processing[J]. Journal of Computer Research and Development, 2009, 46(10): 1634-1641.
    Citation: Tian Hangpei, Gao Deyuan, Fan Xiaoya, and Zhu Yian. Memory Request Queue of Multi-Core Multi-Threading Processor for Real-Time Stream Processing[J]. Journal of Computer Research and Development, 2009, 46(10): 1634-1641.

    面向实时流处理的多核多线程处理器访存队列

    Memory Request Queue of Multi-Core Multi-Threading Processor for Real-Time Stream Processing

    • 摘要: 针对多核多线程处理器中乱序访存影响计算实时性的问题,在对典型访存队列进行研究的基础上提出了一种新的访存队列构建模型及其硬件结构.该模型采用窗口优化算法控制最差情况下的访存延迟,保证访存的实时性,同时又利用优化的乱序调度策略减少访存延迟.实验证明,该访存队列可控制最大访存延迟,与顺序访存相比,存储器具备更高的带宽,与传统的乱序访存相比较,可以充分满足计算的实时性需求,而存储器有效带宽基本不受影响,解决了多核多线程处理器承担实时流计算的基础难题.

       

      Abstract: In order to improve the bandwidth of DRAM, memory request queue inside the memory controller usually equips out-of-order scheduler, which affects the real-time stream processing of multi-core multi-threading processor. The authors propose a novel memory request queue to solve the problem based on studying typical structure of memory request queue. The scheduler of the new memory request queue dispatches memory request with out-of-order scheduler, but is controlled by window. Window defines the certain number of memory requests which are visible to scheduler. The scheduler receives memory request of new window only when finishing all of the requests belonging to the current window served. On the one hand, the out-of-order scheduling algorithm based on parallel execution can improve the memory bandwidth effectively, and on the other hand, the window constrains the largest memory request delay occurring in the worst case by preventing the scheduler from postponing the execution of memory request for unlimited duration. The simulation shows that the memory request queue proposed can adjust the largest memory request delay with scalability and the hardware cost is small. Compared with general out-of-order memory request queue, this presentation not only retains the memory bandwidth, but also supports real-time stream processing, thus proposing a new way of real-time computing with multi-core multi-threading processor.

       

    /

    返回文章
    返回