The High Time Efficiency Task Assignment Strategies For Heterogeneous Computing Systems

With the increasing complexity of tasks in computing systems,the data to be processed has the characteristics of big computations,various types and high complexity,and traditional homogeneous computing systems can no longer satisfy the data processing requirements.To improve the computing power,the computing systems have gradually evolved from homogeneous processors to heterogeneous processors,from centralized systems to distributed systems,from heterogeneous system on chip to heterogeneous distributed systems.To run multi-tasks efficiently and reliably in heterogeneous computing systems,task assignment strategy has become an urgent problem to be solved.In this paper,we study the efficient task assignment of CPU-FPGA-based heterogeneous system on chip,small-scale embedded heterogeneous distributed computing system and large-scale service-oriented heterogeneous distributed computing system.The main contributions of this paper are as follows:1.According to the acceleration requirements of convolutional neural network inferences in CPU-FPGA-based heterogeneous system on chip.We first introduce the high performance hardware accelerator called Deep Learning Processing Unit(DPU),and evaluate the main causes of low schedule efficiency of DPU-based system.Then,we explore the scheduler rules and improve its schedule efficiency through observations and analysis.Finally,we integrate our optimization,and propose an efficient task assignment framework to maximize performance.Experimental results on Xilinx Zynq Ultra Scale+ MPSo C zcu104 show that our efficient task assignment framework significantly boosts schedule efficiency for small-scale networks(from36% to 70%),medium-scale networks(from 65% to 95%),and large-scale networks(from 77% to 99%)compared with original schedule strategy.2.According to the requirements of low redundancy and high time efficiency in embedded heterogeneous distributed systems,the fast task assignment for optimal redundancy(FTAOR)algorithm is proposed to satisfy the application reliability requirement.Unlike exhaustive method,the verification time is reduced by using the FTAOR algorithm,which predicts and eliminates most meaningless computations effectively.Experimental results on parallel applications at different heterogeneity show that the proposed FTAOR algorithm generates lower redundancy compared with the state-of-the-art heuristic algorithms and increases the task size from 30 to 150 with a shorter computation time compared with exhaustive method.3.According to the requirements of large scale tasks and high time efficiency in service-oriented heterogeneous distributed systems,the fast task assignment for minimizing redundancy(FTAMR)algorithm is proposed to satisfy the application reliability requirement.The FTAMR algorithm fast identifies tasks which need to be replicated,and fast maps selected tasks to their respective most suitable processors.Then,the FTAMR algorithm repeats above steps until application reliability satisfies established reliability requirement.Experimental results on real and synthetic generated parallel applications at different scales,parallelism,and heterogeneity show that the FTAMR algorithm can generate minimum redundancy and maximum time efficiency compared with the state-of-the-art fault-tolerance algorithms.