Research On Performance Optimization Of Three-dimensional Network-on-chip For Multicore Processors

In order to meet the growing performance requirements,the number of processor cores continues to increase.However,when the number of cores increases to a certain level,the traditional bus interconnection structure faces great challenges in communication performance,power consumption and expansibility.The communication between cores of multi-core processor has become the bottleneck of the overall performance improvement of multi-core processor system,and Network-on-Chip(No C)emerged.As a new structure,ThreeDimensional Network-on-Chip combines the advantages of Three-Dimensional Integrated Circuit(3D IC)and No C.Compared with the traditional No C,the performance of 3D No C is significantly improved,and has become a research hotspot in the field of No C.In multi-core processors,the interconnection network,as the communication backbone between processor cores,plays a decisive role in the overall performance of the system.In order to improve the overall performance of multi-core processors,the research on performance optimization of 3D No C is very important.3D No C topology reflects the layout and connection of communication nodes in the chip,and greatly influences network performance.This dissertation mainly studies the design and optimization of 3D No C topology under specific applications,so as to optimize the performance of 3D No C and improve the overall performance of multi-core processors.The innovation and main work of this dissertation are as follows:For the performance optimization of 3D No C,this dissertation models the performance optimization of 3D No C and establishes the overall framework of performance optimization.On this basis,this dissertation implements the 3D No C performance optimization algorithm based on Simulated Annealing(SA algorithm)and the 3D No C performance optimization algorithm based on Machine Learning(ML algorithm).In order to make a trade-off between different application objectives of 3D No C performance optimization,the multi-objective optimization methods are introduced into ML algorithm and a 3D No C performance optimization algorithm based on Machine Learning under multiple application objectives(MA-ML algorithm)is designed and implemented in this dissertation.For the performance simulation of irregular topology,this dissertation studies the routing algorithm for irregular topology.This dissertation implements Up*/Down* routing algorithm on gem5 simulator,and improves the performance of Up*/Down* routing algorithm by using three methods: spanning tree optimization,root node selection optimization and traffic balance optimization.To enable more nodes in the network to transmit data over the shortest path,this dissertation also implements the Up*/Down* Escape routing algorithm based on the escape channel.Finally,this dissertation conducts corresponding experiments on the above optimization algorithms and routing algorithms.Firstly,based on three optimization algorithms,3D No C performance optimization experiments are carried out for different application objectives to generate 3D No C design.Then,based on the 3D No C design generated by ML algorithm,simulation experiments are carried out under different routing algorithms to compare the performance of routing algorithms.Finally,based on the routing algorithm with the best performance,the 3D No C design generated by three optimization algorithms is simulated to compare the performance of the optimization algorithm.The experimental results show that in terms of efficiency,compared with SA algorithm,ML algorithm increases by 133% and MA-ML increases by 24.7%.In single application,the average hops and delays of 3D No C optimized by ML algorithm are reduced by 29.1 % and12.4% compared with 3D Mesh and 7.2% and 4.2% compared with SA algorithm.Under multiple applications,the 3D No C optimized by MA-ML algorithm is reduced by 31.6% and14.6% compared with 3D Mesh,and 7.3% and 4.2% compared with ML algorithm.To sum up,ML algorithm is superior to SA algorithm in both optimization efficiency and optimization performance.MA-ML algorithm designed and implemented in this dissertation is slightly insufficient compared with ML algorithm in optimization efficiency.However,through the introduction of multi-objective optimization method,the performance of 3D No C optimized by MA-ML algorithm is improved compared with ML algorithm under multi-application,and the 3D No C optimized by MA-ML algorithm can better adapt to multi-application scenarios.Under multiple application objectives,the MA-ML algorithm proposed in this dissertation can effectively improve the performance of 3D No C,which has practical value and application prospect.