Research On Energy-efficient Task Scheduling And Resource Reconfiguration In Universal Heterogeneous Computing Platform

With the development and progress of computer science technology and microelectronics science,requirements for high-performance data processing are presented in the era of digitalization and information.Besides,traditional general homogeneous computing systems are faced with the bottleneck of data processing.Based on the above two points,general heterogeneous processing systems with more powerful capability and rich processing resources are proposed and extensively researched.With the increase in computing power,the energy consumption of processing system becomes a hot issue in the field of high performance computing.Starting from the general heterogeneous processing system and aiming at the key issues of system architecture,component model,reconfigurable design and resource scheduling,this paper systematically expounds relevant theories and analyzes the technical route.Finally,solutions for the above issues are proposed and performance and usability are analyzed.This thesis studies the hardware and software composition of the general heterogeneous processing system.Based on the hierarchical design idea,the general heterogeneous processing system is divided into user layer,application service layer,middleware layer,device driver layer and heterogeneous hardware layer to build a system model with decoupling and layering of software and hardware.Simultaneously,the software and hardware data interaction method under the general heterogeneous processing system is studied,and the software component model is designed to meet the requirements of resource reconfiguration.A common data format for software platforms is designed to describe configuration files.The software and hardware communication message format is designed to describe the content of components conveniently and avoid subcontracting and sticking.Finally,four reconfigurable FPGA methods are designed based on the idea of combining software and hardware.On-board verification and simulation verification show that vari ous reconfigurable methods are able to obtain good feasibility.Considering the importance of system energy consumption to general heterogeneous processing systems,this thesis proposes two energy-saving task scheduling methods with the goal of optimizing system energy consumption under the premise of not affecting the overall completion time of task scheduling.Based on the dynamic voltage frequency scaling(DVFS)method in the relaxation time,the first task scheduling method treats multiple tasks with associated relaxation time as a group of tasks and adjust the processing frequency of computing nodes to keep the computationally expensive tasks in the associated tasks in a low frequency state for as long as possible to save the system energy consumption.The tasks with large computational overhead are kept in a low frequency state for as long as possible,so as to achieve the purpose of saving system energy consumption.Experiments show that this method can save up to 8% of energy consumption compared with the heterogeneous earliest finish time algorithm.The second method is based on the theory that tasks with long service time in the communication network will prolong the overall completion time of the system and a communication cost sensitive factor is designed.By shutting down computing nodes with high average computational cost in the system during iterations,the system energy consumption is optimized.Experiments show that the proposed scheduling algorithm can achieve up to 11% energy saving ratio.By using the proposed methods,the energy consumption ratio and relative scheduling length ratio can be reduced by 12% and 9% respectively.