Design And Implement Of A Dataflow Programming Model On The Sunway TaihuLight

The integration of massive heterogeneous computing resources within one chip have been a trend of modern many-core processor design.The Sunway TaihuLight,the supercomputer with the highest floating-point performance in the world so far,is made up of 40960 heterogeneous many-core processor SW26010,and each SW26010 has 260 computing cores.Limited to the describing ability of the programming seman-tics of existing parallel programming models(such as MPI and athread)of the Sunway TaihuLight,how to use these programming models to generate enough computing tasks for numerous computing cores?balance workload among heterogeneous computing re-sources and handle data dependencies between computation on the Sunway TaihuLight,brings great challenges to users.The Codelet model is a fine-grained program execution model that combines data flow method and control flow method.It is proposed by Professor Guangrong Gao of University of Delaware.Based on the Codelet model,this paper designs a dataflow-based programming model,SunwayFlow,and implements its runtime system on the Sunway TaihuLight,which aims to provide a way for programmers to effectively solve their computing problems and efficiently use the computing resources on chip.In this paper,the benchmark,HPCG,is used as our case study to verify the effectiveness of SunwayFlow dataflow programming model and its runtime system.The main contributions of this paper are as follows:(1)Design a dataflow-based programming model,SunwayFlow,on the Sunway TaihuLight.Based on the Codelet program execution model,we propose a scheme of mapping abstract machine model onto the Sunway TaihuLight,remove the TP structure,introduce the prefired state,and define the programming interface of the programming model.(2)Implement a runtime system of SunwayFlow model on the Sunway TaihuLight.Design the execution mechanism of the runtime system according to the characteristics of hardware and software of the Sunway TaihuLight,and improve its performance and usability by several key technologies.Users build their own dataflow program by using the programming interface of the runtime system.And then The runtime system can efficiently and automatically accomplish the scheduling and computing of the program.(3)Refactor all computing kernels of HPCG through SunwayFlow model.Es-pecially for parallelizing the symmetric Gauss-Seidel relaxation(SymGS)algorithm,we introduce a novel parallel method,the dataflow method.In the test of single core group,compared to the HPCG reference implementation,our optimized implementa-tion achieves a speedup of 11.79X and 10.31X in SymGS executing time and HPCG performance respectively.