Researches On Heterogeneous Computing System And Task Scheduling Algorithm Based On CPU-GPU-FPGA

With the advancement of computer science,"Big Data" is in the ascendant,and the Internet of Things business is springing up.The amount of data and computing needs in various fields are climbing.In this era of "big data" requiring "big computing",the tradition CPU-based computing systems are incapable,and quantum computing and neural network computing with subversive changes are not available in practical applications.Therefore,in the field of high-performance computing,heterogeneous computing has a tendency to "taking the lead in the world".Research on heterogeneous computing systems is related to national economy and people's livelihood.Tracing back to the source,the reason of heterogeneous computing shining in the field of high-performance computing,on the one hand,it benefits from the rich hardware resources;On the other hand,it is necessary to make each computing unit perform its duties and do its best.Therefore,reasonable task scheduling is indispensable.Task scheduling research can make the computing power of heterogeneous systems go a step further,which is of great significance for heterogeneous computing systems,for the everincreasing computing needs.At present,heterogeneous computing systems are generally equipped with GPU acceleration components.The effect of computing-intensive tasks is obvious.Due to their own architecture,the GPU is deficient in communication-intensive tasks.In addition,the task scheduling algorithms for heterogeneous systems are numerous,but the scheduling length is unsatisfying.In this paper,the following aspects are studied for heterogeneous computing systems and heterogeneous scheduling:A heterogeneous task scheduling(GA-TP)algorithm based on novel genetic and task aggregation is proposed.The global and local algorithms are added by adaptive dynamic selection,whole-genetic hybridization,and adaptive dynamic mutation.The optimization ability enables the GA-TP algorithm to give an infinite approximation optimal scheduling scheme in a reasonable time;The number of tasks and the communication cost between tasks has been reduced by the task aggregation.Further,the scheduling length is shortened.The simulation experiment is designed.The GA-TP algorithm and the traditional algorithm are compared by key parameters such as scheduling length,scheduling length ratio,algorithm acceleration ratio and calculation-communication ratio.The experimental results show that the scheduling length of GA-TP algorithm is significantly shortened.A heterogeneous computing system based on CPU-GPU-FPGA is built,and both GPU and FPGA acceleration units are supported to support computing-intensive,logic-intensive,and communication-intensive computing tasks.Aiming at the delay problem of data transmission between different architecture computing units,combined with the characteristics of the development framework and the characteristics of the GPU,an optimization method is proposed to improve the continuity of time and space of the work-item when them access the global memory.Futher,the optimization method also reduce the GPU's access frequency to the global memory and the clock cycle occupied by the GPU access memory,thereby reducing the data read and write delay.The data transmission link between the GPU and the FPGA is realized.The feasibility of data transmission link is verified by experiments.