Research On Optical Interconnection Techniques For Multiprocessor Network

The expanding gap between microprocessors and interconnection network performance has influenced the development of the HPC (High Performance Computing). As the electrical interconnects is limited due to the cross-talk and losses, electric links will not meet the need of high speed interconnection technology. Optics offers important advantages, it can connecte two nodes with high bandwidth, low power dissipation, little cross-talk and low latency, and optical technology has been used in intra and inter system connections. Although the optics can improve the performance of the interconnection network greatly, it is still not easy to solve the problem such as to have a well-worked topology in optic network and a suitable optic links in board.This paper carries out a comprehensive research on the topology of optic network and investigating the advantages and the disadvantages of different optic bus network. The interconnection network of the multi-processor system has some specialties, and the traditional optic network may not meet its need. In this paper, a new topology called CSAPID (Centralized Scalable All-Photonic Interconnect for Distributed shared memory) is put forward to solve the problem. The CSAPID provides high connectivity by maximizing the channel availability for multiprocessor system interconnection network by using a combination of aggressively designed wavelength, time and space-division multiplexing techniques. This paper evaluates CSAPID based on network characteristics and minimizes the remote memory access latency by reducing the latency between nearbyhood nodes, and experimentalizes the performance of the CSAPID by using OMNeT++. Compare the CSAPID with RAPID, the data transmitted in the net has a lower latency when the net is large. As the design of the optic links on board is another important technology of the optic network, this paper also analyse the losses in the on-board optic links in the network, and research on the methods to calculate the bending loss in OPCB.Experimental results show that our CSAPID network has a good performance by reducing the latency during chip-to-chip interconnects in DSM system in the case that the remote memory access mostly occurs between neiborhood nodes. The analysis of the optic link loss is also been proved by experiment.