Research On Dataflow Runtime Based On GPU/CPU Heterogeneous Environment

The failure of Moore’s Law has prompted processors to deepen in the direction of multi-core and multicore heterogeneity as the semiconductor industry evolves.How to effectively exploit the potential processing power of heterogeneous platforms is the aim of many parallel computing models,and dataflow models,as part of parallel computing models,are widely used in the fields of big data streaming,molecular prediction and machine learning.The dataflow runtime system plays a key role in the execution of streaming applications as an intermediate layer system between the dataflow model and the heterogeneous platform.The main contributions and research of this thesis are as follows:1.The overall design of the data flow runtime system is proposed.The runtime system design is divided into design goals,functional design and architectural design by analysing the execution flow of the heterogeneous GPU software systems CUDA and Open CL with respect to the overall implementation details of the runtime system.In the design objectives,the initial model of the object under study is given,and the system functions are split into a data flow task module,a dependency management module,a data flow graph module and a data flow task scheduling module.In the architectural design,the business scope and interrelationships of each functional module are described in detail,and the proposed overall design plan provides a business orientation for the subsequent implementation of a real runtime system.2.Simulation implementation of dataflow task scheduling system in heterogeneous multi-core GPU environment.The simulation implementation of the heterogeneous environment provides executable support for the real hardware system.In view of the fact that the abstract program model of the dataflow scheduling system is prone to unsafe data flow legends,the thread-safe random generation algorithm proposed in this paper generates graph datasets through the principle of sequential consistency.,which can effectively prevent the dataflow graph from forming a closed-loop structure.In addition,in the face of the complexity and low scheduling efficiency of the simulation scheduling process,the execution model designed in this paper simplifies the abstract machine model and thread-level execution process.Finally,a load-aware queue division scheduling algorithm is proposed,which further analyzes the influence of predecessor and successor and communication overhead on the scheduling system,improves the priority calculation method,and determines the priority of tasks by the extreme value of front-sequence priority,communication overhead,and calculation overhead.In the server selection stage,the scheduling queue is selected optimally by locality.3.Implementation of a dataflow runtime system in a heterogeneous multi-core CPU environment.Based on the simulation system design,this thesis modularizes the dataflow runtime system into four parts,dataflow task module,dataflow dependency management module,dataflow graph module and dataflow task scheduling module.Facing the problem of poor adaptability of dataflow tasks,the task module expands dynamic and combination tasks to improve the scope of applicability of graphs.For dataflow graphs of different densities,which are prone to difficult management and poor scalability,an implicit dependency declaration method is introduced to determine tasks Dependencies between graphs reduce the complexity of graph definition.The dataflow graph module designs graph storage structures,monitors,and executor components and implements memory optimization algorithms for compound tasks to improve the memory utilization of dataflow graphs.Finally,improve the original work-stealing scheduling algorithm and use task locality to speed up the execution speed of stream programs.