Large-Scale Stream Processing Task Resource Scheduling Method Based On Deep Reinforcement Learning

At present,the processing logic in the distributed stream processing system has become more complex,the scale of the data flow graph has become larger,and the graph topology has become more complex.The heuristic scheduling rules in the existing stream processing resource managers are not flexible enough,and the scheduling effect is poor,which affects the performance of the stream processing system.The existing node scheduling algorithms have facing challenges in achieving ideal scheduling effects on large-scale stream processing tasks.The problem of how to schedule the nodes of large-scale data flow graph in stream processing applications and improve the processing efficiency of distributed stream processing systems is discussed.The resource scheduling method of the large-scale stream processing task is designed and implemented based on deep reinforcement learning and hierarchical graph partition technology in this reseach.A method of stream processing resource scheduling based on reinforcement learning is proposed.A data flow graph encoder model based on graph embedding and a graph-aware device allocation decoder model are constructed,and jointly constituted a graph-aware encoder decoder model.The encoder decoder model calculate the equipment resource allocation result in stream processing system for nodes after trained via a reinforcement learning method.A large-scale stream processing task resource scheduling method based on the idea of multi-level graph partitioning is further proposed.In this method,the graph coarsening method is used to reduce the node size of the data flow graph in order to convert the problem into a scheduling strategy learning problem on a simple data flow graph,reduce the difficulty of problem processing and limit the problem solving space.Finally the obtained solution base on a small problem scale is mapped back to the solution of the original problem.In this study,two resource scheduling methods for stream processing tasks are proposed,which have achieved better scheduling results in large-scale stream processing task data sets,and provide a feasible solution for large-scale stream processing task resource scheduling problems,and at the same time prove the effectiveness of graph coarsening process the in optimizing the resource scheduling effect of large-scale stream processing tasks.