异步光分组交换型高性能计算机系统架构及性能分析

提出基于异步光分组交换及组播技术的新型高性能计算机系统架构。采用分布式管理的树型结构，可实现全光多级扩展。采用3.2Tbit/s的超大容量光纤，单信道容量为40Gbit/s，两级系统可以实现6400个CPUs 的互连。采用分组重传冲突解决机制解决系统内部冲突，并仿真了系统的稳定性要求及时延特性。系统稳定性随旧包发送速率b及旧包个数k的增大而增大，当b增至无穷大时，分组到达率λ的最大取值趋于800packets/s。冲突导致的分组等待时延随λ的增大而增大，随服务窗口数C及分组离开率μ的增大而减小，当C增加到6个左右，或者μ增大至1000packets/s时，分组等待时延减小为10-10s量级。多级系统的稳定性条件与两级系统相同。

Architecture and performance analysis of the novel high performance computing system based on asynchronous optical packet switching

A novel high performance computing system (HPCS) based on asynchronous OPS and optical multicast is presented. Distributed management is realized in the tree-like structure and can easily achieve all-optical scalability. A 3.2Tbit/s huge capacity fiber with 40 GBit/s single channel capacity is used, and it can interconnect 6400 nodes in two-stage architecture. The packets retransmission mechanism is used to resolve the system collision. Stability and packets waiting delay of the system are simulated. System stability increases with the number (k) and sending rate (b) of old packets, and when b increases to infinity, the maximum value of packets arriving rate (λ) goes to 800packets/s. The packets waiting delay induced by the collision increases with λ while decreasing with the packets leaving rate (μ) and the number of the service windows (C). When C is more than 6 or μ increases to 1000packets/s, the waiting delay decreases to the order of 10-10 seconds.