Research On Design Method Of Reconfigurable Hybrid Optical-electronic Networks-on-chip Based On Accelerated Network

With the increasing scale of multi-core processor systems,the development of traditional electronic network-on-chip is limited by the performance of delay,bandwidth and power consumption.Hybrid optical-electrical network-on-chip has the advantages of low communication delay and high transmission bandwidth.It is an effective solution for large-scale multi-core system-on-chip communication.Optical links in hybrid network can accelerate the transmission of global communications,but also bring huge static power overhead.Moreover,the spatial and temporal distribution of global communications is different.Fixed optical interconnection bandwidth supply will lead to inadequate utilization of communication links,and the overall energy efficiency of hybrid systems will be inferior to electronic network.Therefore,it is of great significance to reduce the static power consumption of optical network and to study reconfigurable hybrid optical-electrical network-on-chip.Based on the analysis of optical and electrical network-on-chip topology,and on the basis of optical bus and Mesh architecture,a hybrid 3D optical-electrical interconnection architecture based on accelerated network is proposed,which provides low delay,high bandwidth optical network and high flexibility,low cost electrical network for global and local communication respectively.In order to realize the topological reconfiguration of optical nodes,a mathematical model of the optical node layout is established according to the principle of reconfiguration.Based on the integer linear programming method,the optical node layout is optimized.According to the optimization results,the reconfigurable hybrid optical-electrical network topology RHOE-NoC is designed.In order to make full use of communication resources,a dynamic reconfiguration algorithm of node granularity based on machine learning is designed.By calculating the link utilization and buffer occupancy,and combining with the network load prediction value,the reconfiguration of optical nodes is determined,and the communication bandwidth is dynamically adjusted,which effectively reduces the power consumption.In order to solve the influence of the dynamic change of the topology on data communication,the interactive mechanism of the information of the topology reconfiguration is analyzed,and an adaptive routing algorithm based on local topology perception is proposed.When the optical node reconfiguration occurs,alternative optical nodes in the communication domain are dynamically selected for global communication,which effectively reduces the routing decision-making time and improves the network communication performance.In this paper,the whole system heterogeneous multi-core simulator JADE is modified,the photoelectric interface conversion unit is added,the photoelectric device power calculation model is embedded,and the simulation environment supporting the photoelectric hybrid chip network is constructed to complete the performance verification of the proposed RHOE-NoC.When the network is close to saturation at 8×8 topology scale,the communication delay of RHOE-NoC is 50.6% lower than that of EMesh,the throughput is increased by 14.3%,and the power consumption is reduced by 25.9%.Compared with the routing algorithm based on global topology information table,the local topology-aware adaptive routing algorithm reduces the communication delay by 69.5%,improves the throughput by 33.3% and reduces the power consumption by 24.2% when the network is near saturation.