Research And Design Of Optical Interconnection Network For Hybrid Main Memory System

In recent years,with the rapid development of information technologies such as artificial intelligence,cloud computing and big data,the data scale showed explosive growth.Hybrid memory system,consisting of both dynamic random access memory(DRAM)and non-volatile memory(NVM),has great advantages to single ones,which makes hybrid memory system promising solutions to resolve technical bottleneck and performance challenges experienced by DRAM-only and NVM-only memory systems.Therefore,it has become a popular design for computer architecture in order to achieve the memory system with large capacity,low cost and high scalability.However,the existing hybrid memory system mostly uses an electronic bus-based interconnection network to accomplish memory access between the processor and the memory.As the number of processing cores and heterogeneous data-intensive applications on the chip continue to increase,the memory access will become much more,which will lead to the electronic bus-based network to become the vital performance bottleneck for memory system because of its low parallelism,poor scalability and high power consumption.In order to solve the above problems,the main research work of this paper is to use the optical interconnects with high bandwidth density,low latency,high energy efficiency and bit rate transparency to replace the conventional electronic interconnects.Then according to the communication characteristic of hybrid memory system,design an optical interconnection network for hybrid memory system with high parallelism,high scalability and high energy efficiency.According to the memory level where the DRAM and the NVM are located,the existing hybrid memory system can be divided into the traditional hybrid memory system and the hierarchical hybrid memory system.The design of optical interconnection networks for them are different due to the different memory access characteristics.Firstly,we investigate the communication characteristics of same-level hybrid memory system,and therefore design a scalable optical interconnection network topology.Furthermore,we propose an optimized strategy of wavelength allocation based on the single Latin Matrix.This wavelength assignment enables optical interconnection network to support non-blocking parallel communication between the processing element clusters and memory modules with lower wavelengths and optical devices.Then,the scaling rules of optical interconnection network topology and wavelength allocation at different scales are introduced in detail.The proposed optical interconnection network and traditional electronic bus-based interconnection network(baseline)are simulated by NVMain,which is dedicated to simulate the traditional hybrid memory system.The simulation results show that compared with baseline,the total max amount of transactions completed in the same simulation cycle increased by 71.30% in the proposed optical network with 4 parallel memory modules and 163.19% in the proposed optical network with 8 parallel memory modules under PARSEC benchmark suite.The bandwidth increased by at least 24.56% and 51.49%;the averaged latency maximum decreased by 40.77% and 58.16%,respectively;energy efficiency improved by 24.54% and 41.45%.This paper further investigate the communication characteristics of hierarchical hybrid memory system,which differs from the traditional hybrid memory system.An optical interconnection network topology for hierarchical hybrid memory system is proposed in this paper,which urges re-construction of the communication path in network topology to achieve the non-blocking parallel communication between the processing element clusters and DRAM cache modules,the DRAM cache module and NVM memory module.According to the multi-wavelength routing method in the network topology,an optimal wavelength allocation based on the multiple Latin Matrices is proposed.Then,the scaling rules of optical interconnection network topology and wavelength assignment at different scales are introduced.The simulation results show that the proposed optical interconnect network for hierarchical hybrid memory system significantly improves performance of total transactions,bandwidth,average latency and energy efficiency compared with the electronic bus-based interconnection network.