| With the rise of computing intensive fields such as artificial intelligence,deep learning and data analysis,application types are gradually diversified,and computing is developing towards complexity and diversification.The general computing represented by CPU is facing many problems,such as insufficient bandwidth,high energy consumption and high delay,and gradually loses its original advantages.Rely on special hardware design,GPU and FPGA can realize almost all kinds of applications efficiently and flexibly,which greatly improves the computing performance and resource utilization of computing cluster.Heterogeneous computing based on "CPU + FPGA" or "CPU + GPU" has gradually come into public view.However,heterogeneous computing does not mean to put all kinds of chips together mechanically.Different chips have different processing performance for tasks with diverse computing features.In cluster,how to assign tasks to heterogeneous devices is a NP-hard problem.This paper proposes a heterogeneous device scheduling scheme based on "CPU + GPU + FPGA" heterogeneous platform.Specifically,through the establishment of task model and heterogeneous device evaluation model,a two-tier scheduling mechanism is designed to map computing tasks and heterogeneous devices.In order to support preemptive scheduling,this paper introduces checkpoint restart mechanism in the computing platform as well,and inserts checkpoints into tasks to improve the flexibility of task scheduling.Finally,multiple test cases are selected to design functional experiment,heterogeneous computing performance experiment and heterogeneous scheduling performance experiment.The results show that the platform is practical and has the effect of accelerating general programs.At the same time,the heterogeneous scheduling scheme improves the computing performance of the platform,realizes the task to be processed by multiple heterogeneous devices,and basically achieves high efficiency,low latency and low power consumption. |
PDF Full Text Request