Research On Exa-scale HPC Data Center Optical Interconnecting Network

Research institutions predict that high-performance computing?HPC?systems will achieve the Exa-scale around 2020,namely the computing capacity will reach 10¹⁸ Flops.As a key component of a HPC system,the interconnect network bears a huge amount of data generated by various applications.Because different applications have different traffic patterns,and different topologies are suitable for different traffic patterns,matching the topological structure with the traffic patterns of the application can improve the communication performance and resource utilization of the network and reduce costs and energy consumption,and improve network flexibility.At the same time,optical interconnect technology has become one of the solutions for high-speed interconnect networks in HPC systems due to its advantages of high bandwidth,low power consumption,high reliability,and high flexibility.The choice of exchange mechanism and optical switching device in the optical interconnect network is more flexible.Selecting the switching mechanism and switching device that match the application traffic patterns can not only overcome the defects of long switching device configuration time,but also improve the throughput of the network.Balance the network's cost and performance,making the network more flexible.This thesis aims at the traditional optical/electronic hybrid network architecture based on tree topology,which is difficult to meet the multi-dimensional,continuous and stable traffic patterns in HPC applications,and the demand for continuous circular subnetworks during task allocation.An application-driven optical/electronic hybrid interconnect network TD-HyEON is proposed.The top layer of the TD-HyEON network architecture adopts an optical interconnection network based on a micro-motor system?MEMS?optical switch.The optical network uses an optical circuit switching?OCS?mechanism to provide high-bandwidth and reliable communication light for sustained and stable traffic.TD-HyEON's bottom network adopts electrical interconnection networks based on commercial electrical switches,and we design a new topology based on Cube and Ring to meet multi-dimensional communication such as colletive and continuous ring subnetworks demands;At the same time,the control layer of the TD-HyEON network achieves the maximum throughput of the optical network through the configuration of MEMS optical switches and improves the utilization of wavelength resources.Simulation results show that TD-HyEON network can effectively improve network communication performance and balance network performance and cost overhead.In this thesis,for the pod-based modular HPC system interconnection network,a single topology in a pod cannot meet the traffic patterns of multiple applications,and a pod-level reconfigurable optical interconnection network ReON is proposed.The ReON uses an array-based waveguide grating router?AWGR?optical interconnect network to interconnect each compute node with an intra-AWGR inside the rack.Inter-AWGRs are used between racks to connect each intra-AWGR.ReON controllers can dynamically adjust the AWGR connection and node communication wavelength to achieve dynamic reconfiguration of the network topology.When the ReON can be reconfigured as a Torus structure and a HyperX structure,we have designed a wavelength allocation algorithm and an AWGR connection relationship matrix calculation algorithm.We use simulation based on real HPC applications test data.Simulation results show that ReON can effectively improve network performance compared to a single topology.