Research On Data Flow Actors Scheduling Based On GPU/CPU Heterogeneous System

With the development of processors tend to be multi-core and specialized,heterogeneous computing platforms represented by multi-core CPU and multi-core GPU are widely used in the fields of model simulation and molecular prediction.In order to make the most of the advantages of heterogeneous computing platforms,it need reasonably allocate computing tasks to computing units.Stream application is one of many computing tasks.It exists in the field of signal processing and machine learning,and can be abstracted as synchronous data flow graph.Synchronous data flow graph shows the parallelism inside the program explicitly.The research on scheduling of synchronous data flow graph is becoming more and more popular.However,the existing synchronous data flow graph scheduling algorithms are difficult to achieve good scheduling results in heterogeneous systems.The main research contents and contributions of this thesis are as follows:A general parallel computing model for data flow SDAG model is proposed.Aiming at the lack of general parallel computing model of data flow program under von Neumann architecture,this thesis takes synchronous data flow graph as the research object,and studies the abstract representation of synchronous data flow graph-directed acyclic graph and data flow computing idea.Based on the directed acyclic graph model and the idea of data-driven,a SDAG parallel computing model is proposed.In SDAG,communication and data driving are realized by designing specific queues.SDAG model indicates the creation process and communication method of data flow actor.The list scheduling algorithm is improved.In order to verify the performance of scheduling algorithm and train scheduling agent,a random generation algorithm of synchronous data flow graph is designed and implemented to generate synchronous data flow graph to meet the needs of users.In view of the problems of computational complexity,low parallel utilization and load imbalance in the existing scheduling algorithms,this thesis first improves the priority calculation method to avoid creating additional vertices in the original graph,and reduces the time complexity of priority calculation.The short job priority strategy is used in the instance selection stage and the earliest completion time strategy is used in the processor selection stage,which improves the parallel utilization of the scheduling algorithm and the load balance of computing resources.A scheduling method based on deep reinforcement learning is proposed.Aiming at the bottleneck problems such as poor flexibility of heuristic scheduling strategy,this thesis proposes a synchronous data flow graph scheduling method based on deep reinforcement learning-Milk T.The state space and action space suitable for scheduling synchronous data flow graph are designed,and the representation of graph neural network learning eigenvector is applied to improve the convergence speed of machine learning model.Milk T provides a more potential solution for scheduling synchronous data flow graph actors.In the experimental part,the test data are generated by using the designed synchronous data flow graph random generator,and compared with the existing scheduling algorithms,the average acceleration ratio of the improved scheduling algorithm and the deep reinforcement learning method in this thesis is increased by 63.20% and 22.14%,respectively,under the simulated heterogeneous system.Both methods achieve good scheduling effects,and also prove the effectiveness and flexibility of scheduling synchronous data flow graph actors based on deep reinforcement learning.