Research On Scheduling Algorithm Of Heterogeneous Computing Platform Based On Reinforcement Learning

Heterogeneous computing platform is widely used in cloud computing,data center,Internet of things and other fields because of its rich computing resources,flexible architecture,wide range of applications and strong parallel processing ability.However,its complex system composition brings new challenges to task scheduling.The platform needs to be considered the system architecture,resource attributes,task types,platform constraints and scheduling objectives in order to obtain the ideal scheduling results.The development of artificial in-telligence provides a new way to solve this problem,but the current intelligent scheduling algorithms ignore the heterogeneous attributes of resources,and lack of detailed descrip-tion of task dependency,which leads to poor scheduling effect.In order to solve the above problems,this paper conducts research from two aspects of task and resource modeling,and multi -objective constraint scheduling algorithm.The specific work is as follows:(1)Considering the problems of task data parallelism,highly heterogeneous computing re-sources,task resource type matching and system operation quality that must be solved in the scheduling of heterogeneous computing platforms,this paper designs execution time,power consumption and reliability models respectively.It also accurately characterizes the time ex-penditure,power consumption and reliability of tasks under different scheduling strategies,and provides accurate model support for the design of efficient management scheduling al-gorithms based on comprehensive considerations such as user experience,service provider costs,and platform robustness.(2)In order to solve the problems of large solution space and poor algorithm adaptability of the scheduling strategy in heterogeneous computing platforms,this paper proposes a re-inforcement learning scheduling algorithm based on the deep deterministic strategy gradi-ent algorithm according to the above-mentioned multi-objective optimization model.First,graph embedding technology is used to design the coding of diversified and complex task states,which solves the problem that tasks cannot be fixedly coded in heterogeneous com-puting platforms.Secondly,task embedding module and device embedding module,based on graph neural network,which are responsible for selecting ready subtasks and encoding the set of computing resources? Finally,the Gumble Softmax function is used to solve the non -differentiable problem in the discrete distribution sampling process.Through the design of a scheduling decision module based on the fully connected network,the mapping of sub-tasks to specific physical devices is realized.The algorithms proposed in this paper are compared with Random scheduling,FIFO schedul-ing,SJF scheduling,Roulette scheduling,and existing reinforcement learning scheduling al-gorithms to analyze the performance of different task resource allocations in the system.In the case of changes in the number of cluster servers and the number of tasks on the com-puting platform,taking the average of the average task completion time,the algorithm in this paper is superior to the above-mentioned comparison algorithm in the optimization goal of minimizing the average task completion time.Comparing the multi-objective algorithm which introduces power consumption and reliability constraints with the single-objective al-gorithm that only introduces time constraints,the average task completion time does not increase significantly when the power consumption and failure rate are reduced.The exper-imental results show that the algorithm in this paper can optimize the system reliability and power consumption while reducing the average task completion time as much as possible.