Research On Energy And Delay Oriented Network-on-Chip Mapping Technology

Network-on-Chip (NoC) represents an important solution to on-chip communication ofcomplex MPSoC, NoC Mapping is a significant step in NoC design. Different mapping resultshave a great effect on the energy consumption of the system, delay, throughput, thermal balanceand other properties. At present, NoC mapping studies mainly focus on the constructed delaymodel, performance assessment model， multi-objective mapping algorithm, dynamic mappingstrategy and3D NoC design. With the development of NoC mapping technology, the traditionalsingle performance optimization algorithm may lead to a sharp decline in another performance,while the need of energy consumption and delay oriented multi-objective performanceoptimization becomes more and more obvious. Therefore, the dissertation attempts to researchfrom the model construction and multi-objective mapping algorithm design on NoC mapping.Based on the analysis of existing NoC mapping technology, this dissertation first presents adelay optimization model based on macro link load distribution and single node queuing delay,and a mapping scheme based on nearest neighbor random GA algorithm. And then a mappingalgorithm based on Boltzmann-NSGAⅡ algorithm is proposed aiming at the multi-objectiveoptimization problem of NoC mapping. Furtherly, a mapping algorithm based on chaos NNIAalgorithm is presented. Main work and contributions of this dissertation are outlined as follows:1. An optimization delay model and a mapping scheme based on nearest neighbor randomGA algorithm are proposed. Considering the problem that is hard to model delay accurately, thisdissertation proposes an optimization delay model based on macro link load distribution andsingle node queuing delay. This model introduces the delay factor concept to optimize the twokey aspects: the macro and the single node. Based on the above optimization delay model, amapping scheme based on nearest neighbor random GA algorithm is proposed to solve theproblem that the mapping process of GA is easy to fall into local optimum and low searchefficiency. This mapping scheme constructs the initial population based on nearest neighborrandom thought, design the genetic operators combined with the characters of NoC mapping.The simulation results indicate that the mapping scheme increase the computing efficiency by anaverage of20%compared with GA algorithm.2. A mapping algorithm based on Boltzmann-NSGAⅡ is proposed according to theshortcoming of NSGAⅡ algorithm solving the multi-objective optimization problem of NoCmapping. This algorithm introduces Boltzmann strategy to construct non-dominated sortingmechanism, designs multi-point crossover operator and random mutation operator to avoid therepeated chromosome and conserve the excellent individual. The algorithm guarantees thepopulation diversity effectively and enhances the local search capability. The simulation resultsshow that: Compared with the improved NSGAⅡ algorithm, the optimal solution sets obtainedby Boltzmann-NSGAII promote the convergence by an average of47.4%, the distribution by anaverage of55.5%. The convergence of the optimal solution set obtained by Boltzmann-NSGA IIalgorithm is better, and the solution distribution is more even. 3. Combining with the demand of the NoC development, a mapping algorithm based onchaos NNIA is proposed for the inherent defects of NSGA Ⅱ algorithm solving highdimensional multi-objective optimization problem. This algorithm realizes the recombinationoperation through multi-point crossover operator, combines with the proportional cloning andthe nearest neighbor based selection of the NNIA algorithm, introduces the chaotic updatemechanism into the mutation operator to promote population diversity. The simulation resultsshow that: Compared with NSGAⅡ algorithm, the optimal solution sets obtained by chaosNNIA mapping algorithm promote the convergence by an average of64.4%, the distribution byan average of46.1%aiming at the high dimension multi-objective optimization problem of NoCmapping.