Research On Optical Interconnection Network Architecture In Supercomputers

With today's high-performance computers being comprised of tens of thousands of processors, the trend is for more processors and more memory to achieve better performance. Consequentially, the interconnection network as a medium for communications between processors and memory in a high-performance computing system (HPCS) is facing great challenges. Traditional electricity interconnection is restricted in bandwidth, interconnection density, clock distortion, energy consumption and other respects. However, as many products of optical devices are ripened into maturity, such as VCSEL, WDM, SOA and so on, optical interconnection is developing full of vitality. Because of remarkable advantage of opitcai interconnection network, for example, high bandwidth, low loss, anti-jamming and available parallelization, it can meet the need of developing HPCS. At present, the fiber-based and high-speed optical devices Opto-electronic hybrid interconnection network has been successfully applied to HPCS, and the trend is to develop All-optical network interconnection.The paper firstly introduced the advantage, recent research and future of optical interconnection network, and then analyzed key technologies of the Data Vortex optical packet switching network, such as topology, synchronization, and the solution to packet conflict.On the basis of research in topology and packet transfer principle of the Data Vortex network, this paper compared how much different topology parameters impact on the system performance through mathematical analysis and simulation by OMNeT++ tools and then optimize the topology of the Data Vortex. Research shows that, when the exchange size of the interconnection network is fixed, the Data Vortex with smaller angle parameter achieves better latency and throughput performance.Based on the analysis and optimization of the Data Vortex and considering the novel optical switching device in CMOS processes, we brought up a new packet switching network model based on ring resonator switchs?DDP. A comparison with the optimized Data Vortex was made and the results showed that the DDP performed better, especially on network throughput.