Study On Optimization And Simulation Of 3D NoC Dynamic Thermal Management Methods

With the constant shrinking of semiconductor process dimensions,more and more transistors are integrated into single chip.Meanwhile,improving the performance of single-core chip by increasing frequency cannot satisfy the requirement of application in performance,actually,the many-core and multi-core in system is an inevitable trends.Under this tendency,2D NoC can no longer fulfill the performance needs of inter-core communication in multi-core processors,as a result,3D NoC has gradually been in a hot research in industry and academic owing to its small area and high bandwidth.The three-dimensional stacking structure results in a great difference in the length of heat dissipation path and the heat dissipation efficiency between the silicon layers.It is an urgent task to guarantee higher performance of 3D NoC as much as possible with controllable temperature,the dynamic thermal management method of 3D NoC system is one of the most effective paths to solve it.This paper focuses on temperature prediction of proactive dynamic thermal management method and thermal-aware routing algorithms for 3D NoC,the main research contents and innovations are as follows:1.This paper first analyzes the Access Noxim simulation platform,and then introduces the various components of it.The operation mechanism and implementation principle of the simulation platform have been studied after that.The follow-up research of this paper is based on this platform.2.The traditional reactive dynamic thermal management method starts the temperature management after the temperature is overheated,the over long delay between obtaining and controlling temperature damages the microprocessor chip permanently.In order to grasp the temperature state in advance,this paper studied the equivalent thermal conductivity model of three-dimensional system fully at the beginning.And then,optimized the static temperature prediction method and promoted a dynamic temperature prediction method which extracting R,C parameters in every sampling period.Experiments show that,compared with the single-cycle prediction results of static methods,the prediction accuracy has increased by 92% at the expense of the computational complexity.In the multi-cycle temperature prediction experiment,the maximum average absolute error of the temperature of a single node and a 2*2 region predicted by the dynamic prediction method are 0.009 ° C and 0.086 ° C respectively,which reveals better prediction timeliness.3.In order to control the network traffic according to the heat dissipation efficiency of the silicon layer better,this paper proposed an ILDR routing algorithm which distributing the silicon layer traffic by judging the inter layer distance from the source node to the destination node.The proposed algorithm can set different interlayer distance thresholds according to different network sizes,and has strong controllability for network traffic allocation.Besides,it has been implemented on the Access Noxim simulation platform and compared with two other routing algorithms.The experimental results show that the ILDR algorithm reduces the average delay by 50.59%～86.27% and increases the throughput by 9.23%～25.89% when the packet rate is 0.025,which improves the network performance of 3D NoC system effectively.Although the total energy consumption increases by 9.29%～34.04%,the energy delay product(EDP)decreased by 37.52%～81.41%,actually,the energy use efficiency is higher.