Experimental Development of a Poly-multi Processing Element for a Highly Parallel Computer

本课题研究了一种用于高并行度计算机的处理单元的体系结构设计，并对处理单元上运行的执行线程的管理方法进行了研究。为此，我们开发了一个可以与上一个课题中开发的并行推理引擎PIE64相连的处理单元硬件，并设计了运行在我们的处理单元管理处理器上的并行管理内核程序。处理单元设计的一个问题是通信延迟，这使得处理器的利用率很低，这在大规模并行计算机中是相当重要的。为了解决远程内存访问的延迟，我们提出了多上下文处理方法。为了解决远程内存访问的延迟，我们提出了一种多上下文处理方法，以解决远程内存访问的延迟问题其中最多4个线程可以在单个处理器中运行，逐个周期共享公共流水线。我们开发了一个推理处理器LSI，它采用了我们的多上下文处理方法。并行管理内核是运行在管理处理器上的控制程序。并行管理内核提供了多种并行管理功能，如负载均衡和进程调度。管理处理器采用高性能通用微处理器，该微处理器采用寄存器窗口机制。在管理处理器上运行并行管理内核需要一种快速的上下文切换机制来快速切换并行管理内核的执行。为此，我们提出了一种新的寄存器窗口机制的管理方法，并成功地证明了该方法的有效性。

This research project investigates the architecture design of a processing element for a highly parallel computer, and also investigates the management method of execution threads running on the processing element.For this purpose, we have developed a processing element hardware which can be connected to Parallel Inference Engine PIE64 which we developed in the preveous project, We have also designd the parallel management kernel program which is running on the management processor of our processing element.One problem on designing a processing element is a communication latency, which makes processor utilization low, and is substantial in a large scale parallel computer, To hade the latency of remote memory access, we have proposed "multi-context processing method", in which up to 4 threads can be running in a single processor sharing the common pipeline cycle by cycle. We have developed a inference processor LSI which adopts our multi-context processing method. We have shown the effectiveness of this method using this LSI chip.Parallel management kernel is a control program running on the management processor. Parallel management kernel provides several parallel management functions such as load balancing and process scheduling. We adopt a high performance general purpose micro processor for the management processor.This micro processor adopts register window mechanism. Running parallel management management kernel on a management processor requires a fast context switching mechanism to switch execution of parallel management kernel quickly. For this purpose, we proposed a novel manegement method of a register window mechanism, and we have successfully shown the effectiveness of our method.