| In recent years,the integration technology realizes the need to integrate multiple processors on a single chip,so the multi-processor utilizeing system on a chip has been developed rapidly.Using and lay outing a large number of transistors is the key of the chip architecture in the performance improvement of interconnection delay,storage bandwidth,power consumption management,throughput and dissipation equilibrium.The traditional way of global interconnect based on bus has brought many problems to the system,therefore,the interconnection mode based on communication design——NoC has been developed rapidly and has become the mainstream trend of communication development.However,although NoC has many advantages,the heat problem caused by the high density processor core integration becomes the main reason to hinder the system's heat dissipation.Therefore,the network faces the inevitable heat dissipation problem.This paper chooses from task management,which includes task assignment and task scheduling,to study the balance of heat dissipation.Compared with other heat equilibrium methods,task allocation and task scheduling have been given topology and task diagram,which makes the change of task management and upgrade more flexible,and has small influence on the structure,so the research on thermal equilibrium is less costly.In order to optimize the network heat dissipation,this paper conducts research and analysis from three aspects:task assignment,task scheduling and verification platform for achieving the goals of the temperature gradient reduction and the peak temperature optimization in task allocation,the number of the migration reduction and the average hop count optimization in task scheduling and the verification of temperature by semi-physical simulation platform.This paper first studied task allocation based on Thermal Resistance and Capacitance model as the foundation.Combine the change of the regional vector distance for the processor core affecting temperature with the genetic algorithm,this paper proposes a Regional MeanTemperature(RMT)based on processor core temperature.This method fully takes into account the influence of the region temperature of the processor core.Then,the processor cores temperature gradient is calculated by using vector distance and the genetic algorithm is used to optimize the initial task allocation.The experimental results show that,the peak temperature reduction rate reached 4.69%,the hot spot rate reached 42.31%,and the temperature gradient decreased to 7.49%compared with the random task allocation strategy.For task scheduling problem of NoC,in this paper,we presents a task scheduling based on the shortest Manhattan distance based on the average temperature of the system and the shortest distance path between the source node and the destination node.The method fully considers the shortest Manhattan path between communication nodes in a core communication graph by using the search algorithm to find the destination node for scheduling task and determining the task scheduling pair with simulate annealing algorithm.Experiment results indicate that,compared with the traditional DTM strategy,for 6*6 mesh,8*8 mesh and 10*10 mesh topology,SMDS experiment scheme in terms of number of migration optimize 22.08%,21.74%and 23.02%,respectively and the average optimization rate for average number of jumps is 24.04%,29.18%and 23.04%.Aiming at the establishment of the network simulation platform on the chip,this paper proposes a FPGA semi-physical NoC temperature evaluation platform.For the different modules of the system,the configuration file is generated through the parameterized control of the software interface,and the data flow generation and network performance simulation of the resource node are realized.At the same time,the FPGA can be reconstructed online,and the corresponding topological structure and routing algorithm are generated.Finally,the system temperature can be evaluated. |
PDF Full Text Request