基于异步光分组交换的高性能计算机系统

Asynchronous optical packet switching based high performance computing system

基于异步光分组交换及光组播技术,提出一种新型高性能计算机系统(HPCS)。采用分布式控制结构进行系统的分级管理,有利于处理大批量突发业务及系统的全光多级扩展。基于循环光纤延时线(Rec-FDL),提出一种紧凑、大容量的异步全光分组冲突解决机制,并与传统的分组重传冲突解决算法(PRA)进行了性能比较;建立了系统的稳定性、丢包率及时延分析模型。仿真结果表明,对于一个能够实现40320个CPU互连的两级系统,Rec-FDL机制的平均分组等待时延TRec-FDL约为0.004ns,而PRA算法的平均分组等待时延Trerans随分组传输距离D增加而增加,在D为10m和20m时,Trerans分别等于0.006ns和0.01ns。因此,对于分布式HPCS而言,Rec-FDL机制具有更优秀的时延特性。

An asynchronous optical-packet-switching and optical-multicast based High Performance Computing System （HPCS） with multistage distributed management features is presented, which can process the massive burst services and realize all-optical scalability. A compact and huge-capacity Recycling-Fiber-Delay-Line （Rec-FDL） based collision resolution mechanism is proposed to resolve the contentions for asynchronous arriving packets, and its performances are compared with the traditional Packet Retransmission Algorithm （PRA）. The system analysis models for the stabilities, packet loss rates and average packet waiting latencies of the PRA （Zor^s） and Rec-FDL （TRec-FDL） schemes are built. The simulations show that, for a two-stage system with 40 320 CPUs, TR^c-FDL is fixed as 0.004 ns, while Tror^s increases with the packet transmission distance D, and equals to 0.006 ns and 0.01ns when D is 10 m and 20 m, respectively. Hence, as the HPCS is typically highly distributed, the Rec-FDL scheme can induce lesser latencies than the PRA.