Research On Task Scheduling Algorithm For Heterogeneous Computing Platform

With the development of computer technologies such as big data and cloud computing,and the diversification of computer application scenarios and computing requirements,today’s computer architecture is developing in the direction of isomerization,and efficient data communication methods and task scheduling between heterogeneous processors Algorithms are the two key points to exploit the performance of heterogeneous computing platforms.Therefore,aiming at the CPU-GPU-FPGA heterogeneous computing platform based on PCI-e bus interconnection,this thesis designs an Open CL-based GPU-FPGA direct interconnection communication method and proposes a solution based on problem decomposition for the task scheduling problem on heterogeneous computing platforms,which improves the overall performance of heterogeneous computing platforms.First of all,this thesis introduces the related technologies and heterogeneous computing platforms used in the research work,and studies and analyzes the communication model of the heterogeneous computing platforms,and designs a communication method of GPU-FPGA direct interconnection based on Open CL.At the same time,the overall design of the Open CL kernel module and PCI-e controller module in the design scheme is carried out;Secondly,the task scheduling problem of heterogeneous computing platforms is modeled and analyzed,and a task scheduling model HCP is proposed.According to the model,the task scheduling problem of heterogeneous computing platforms is decomposed into two subproblems: the task mapping problem of processor type and the task mapping problem of computing equipment.According to the task mapping problem of processor type,this thesis proposed (?)-ACO algorithm based on ant colony optimization algorithm.To address the problem of load imbalance in computing equipment task mapping,this thesis proposes a G-S algorithm based on greedy strategy.The experimental results show that the (?)-ACO+G-S can obtain the approximate optimal overall task scheduling problem of heterogeneous computing platforms;Then,this thesis implements the communication method of GPU-FPGA direct interconnection on the heterogeneous computing platform and the test shows that the GPUFPGA direct interconnection communication method designed in this thesis has better communication efficiency;Finally,the (?)-ACO+G-S proposed in this thesis are verified by the YOLO v3 model.The results show that (?)-ACO+G-S algorithm have good results in comprehensively solving the task scheduling problem of YOLO v3 on heterogeneous computing platforms.