Research On Resource Scheduling Algorithm Based On Regional Dynamic Reconfigurable PIM System

Regional dynamic reconfigurable PIM system is a computing system with regional dynamic reconfigurable capability based on PIM technology.By Time Division Multiplexing,system enables large-scale computing applications to be implemented with limited resources.At the same time,the reconfigurable resources in the system architecture realizes the function of storage and computation as one,enabling simulations to be performed,breaking through the limitation of access bandwidth bottlenecks and offering great promise for applications in data-intensive applications such as big data,machine learning,bio-computing and image processing.Effective resource scheduling can give full play to the regional dynamic reconfiguration capability of the system.By reusing reconfigured resources in a time-sharing manner,system latency can be reduced,parallelism and resource utilization can be improved,and the performance of the system can be comprehensively enhanced.However,there are few researches on the resource scheduling of regional dynamic reconfigurable PIM system,and little efficient resource scheduling algorithms are proposed.Therefore,the resource scheduling problem in regional dynamic reconfigurable PIM system is studied in this thesis.The main work is as follows:(1)In order to represent the time-division multiplexing of reconstructed resources,this thesis introduces a time dimension on the two-dimensional plane of the hardware architecture and establishes a model of resource scheduling problem.This thesis also proposes a binary group representation based on the characteristics of the resource scheduling problem,which can represent the resource scheduling results completely.A method to judge the feasibility of the resource scheduling results by time sequence diagram is also proposed.At the same time,in order to improve the application range of the system,this thesis introduced a multi-object optimization function,which provides accurate model for the resource scheduling algorithm proposed later.(2)This thesis proposes a resource scheduling algorithm of regional dynamic reconfigurable PIM system based on the model and simulated annealing algorithm,and makes two optimizations in the design of the algorithm.In terms of the perturbation method of the solution,the algorithm introduces the perturbation method of deletion insertion and uses an incremental evaluation method to assess the cost of the solution,uses the multi-object optimization function to ensure the superiority of the feasible solution and reduce the time complexity of the algorithm.In terms of the annealing table,the annealing process is divided into two stages: high temperature and low temperature.Rapid cooling in the hightemperature phase reduces the time complexity of the algorithm.Tempering is performed in the low-temperature stage to improve the quality of the final solution by expanding the search space of the solution.In this thesis,the influence of the key parameters of resource scheduling model and resource scheduling algorithm of regional dynamic reconfigurable PIM system on performance is examined,and compared with traditional simulated annealing algorithm,integer programming,step-by-step solution and other solutions.The effectiveness,efficiency and flexibility of the algorithm are experimentally verified.The algorithm can effectively solve the resource scheduling problem of regional dynamic reconfigurable PIM system,and can obtain better solution with less time.The algorithm can also be adapted to different application scenarios and achieve good results by reasonably adjusting the proportion of influence factors in the algorithm’s objective function.