Research On Task Scheduling Based On HSA Platform

Traditional homogeneous multi-core processors have not meet the ever-increasing needs of various upper-layer applications for the increasing level of underlying hardware resources in recent years.Processors are now transitioning from the multi-core processor to the heterogeneous processor.Heterogeneous multi-core processors represented by CPU-GPU(Central Processing Unit-Graphics Processing Unit)architecture have gradually become a research hotspot in this field.In 2012,under the advocacy of Advanced Micro Devices,Inc.(AMD),heterogeneous system architecture(HSA)emerged as the times require.In 2014,the Accelerated Processing Unit(APU)that met the HSA computing framework drew great attention from the academia and industry.However,a perfect task scheduling mechanism for this heterogeneous platform is still lacking up to now.And the upper-layer application cannot take excellent advantage of the underlying hardware resources,resulting in a waste of resources.Therefore,there are research questions to be solved in setting up the heterogeneous multi-core cooperative operating environment for APU processors that is in line with the HSA and achieving optimal deployment of various types of workloads by task scheduling.Classical task scheduling research is generally divided into two types: static scheduling and dynamic scheduling.By using a reasonable scheduling method,the performance of the system can be fully exploited and high execution efficiency is obtained.However,due to high platform correlations in the task scheduling,the differences in the underlying hardware architecture makes those two types of scheduling methods designed for other platforms unavailable directly on the HSA platform.The aim of this study is to develop the performance advantages of the HSA platform.The performance model is introduced into the task scheduling research of the HSA platform for the first time in this paper,and the static and dynamic scheduling methods for the HSA platform are provided.Further more,the scheduling methods are applied to the algorithm optimization of polar glaciology.The main work of this paper can be summarized as the following four aspects:1.Using the development of CPU-GPU heterogeneous computing technology as the main research topic,the hardware and software environment for the development of heterogeneous systems is introduced.Investigate the development situation of theHSA ecological environment,and deeply survey the realization process of HSA key technologies,also analyze the work characteristics of dealing with resource competition and collaboration.2.For the purpose of achieving the optimal task scheduling of APU platform based on HSA framework,the HSA performance model is proposed and the basic concepts and construction process is introduced in detail combined with the architecture principle of single-core processor performance model.The processing core execution time curve is firstly obtained by a set of estimation methods in the model.And then the theoretical optimal distribution ratio is obtained based on the intersection point of the curves.3.Based on the idea of establishing the HSA performance model,static and dynamic task scheduling methods for the HSA platform are proposed.The methods are in experimental verification using typical algorithms.Then the advantages and adaptation of the scheduling methods under different conditions are analyzed.4.Taking the typical algorithm,nonlinear Chirp Scaling algorithm in polar glaciology as the case,the basic thoughts of deploying two task scheduling methods to practical applications are explored.The experimental results show that the static and dynamic task scheduling methods based on the HSA performance model reduce approximately 20% of the execution time.Therefore,the methods are proven effective.