高带宽远程内存结构中的预取研究

高速电路和光互联技术的发展极大地提高了网络的速度与带宽。因而，突破高性能计算机CPU与内存紧耦合的传统结构成为可能，CPU与内存的耦合不再受距离的限制，这必将引起体系结构的变革。文[1]提出DSAG结构——CPU与内存在空间上分离，每个CPU节点上仅留少量内存．将海量内存放在远程统一管理作为内存服务器，CPU节点和内存服务器之间通过高速网络互连。这种新的体系结构带来了更好的共享性和可扩展性，但同时也对我们解决CPU和内存之间的不平衡性问题带来了挑战。为了降低DSAG这种远程内存结构增加的访存时延，我们考虑到CPU正常访存没有充分利用网络的高带宽，因此可以利用剩余的网络带宽来进行远程内存数据的预取。本论文在应用程序执行时记录本地（相对于远程内存）不命中的地址信息，以页对齐分析其中存在的页框流（Page Frame Stream）的统计特征，并提出可基于页框流的预取机制可降低访存延迟、提升系统性能的观点。最后我们采用模拟的方法验证了观点的可行性与正确性，进一步提出了三种预取策略，比较并分析影响预取效果的因素。

The Study on Prefetching of Remote Memory Architecture

High speed electrical and optical interconnection technique brings us high-speed and high-bandwidth network. Thus, we can break through the traditional computer architecture by decoupling memory from CPU. Distance between CPU and memory is no longer restricted, and this will consequentially cause innovation in high performance computer architecture. In paper[1]the authors present DSAG architecture-each CPU node is only attached with a small quantity of memory, while massive memory served as memory server is located away, and they are connected by high-speed network. This architecture provides better shareability and more scalability, but it also challenges us to reduce the gap between processor and memory. To reduce the delay of remote memory access, with abundant network bandwidth, we can use the spare network bandwidth while CPU runs to prefetch data from the remote memory. In this paper, we record and analyze the address missed in local memory access while program runs, and analyze the statistical characteristic of the page frame stream. We propose a prefetching approach based on page frame stream to reduce remote memory access delay and improve the system performance. Finally, we use simulation technique to verify the feasibility and correctness of the prefetching approach, and propose three prefetching policies as well as the factors that affect the prefetching.