基于交叉开关的可伸缩新型互联网络

Future massively parallel processors will contain more than ten thousand high performance microprocessors. To connect this kind of large set of processors, a high bandwidth, low latency, and scalable interconnection network is essential. Combining reverse-graph topology and high speed crossbar switch, this project proposes a novel interconnection network NIN to avoid the latency increase problem in k-ary n-cube when network size grows. The project studies the topology propeties, routing algorithms, and multicast communication algorithms, and designs and implements the crossbar switch device, the key component of NIN. With the new crossbar switch, the delay of point-to-point communication is reduced to 3 microseconds.

未来的超级并行计算机将包含数万或更多的高性能处理器。为了将这些处理器有机地连接起来，需要一种高带宽、低延迟、易扩展、可伸缩的高性能互连网络。本项目将反图拓扑结构与高速交叉开关相结合，提出一种新型互连网络结构，以避免目前常用的基于Ｋ元Ｎ方体的互连网络在规模扩大后延迟明显增大的问题，内容包括拓扑性质、路由算法、实现方案等。