A High-Performance Distribution Method of Astronomical Data Based on MPI+CUDA

The appearance of massive astronomical data has brought a lot of challenges to the development of astronomy software. In recent years, with the development of parallel computing technology, MPI+GPU mode has become the main mode for current high performance astronomical data processing gradually. For the problem of how to improve reconstruction performance in reconstruction of solar high resolution image, this paper has made systematic research on data reading and data distribution method. During traditional MPI parallel processing, master process cuts original image into sub-blocks, and then delivers sub-blocks into each sub-process for reconstruction, the results after reconstruction will be returned to the master process. When the number of sub-process is big and calculation nodes are few, this data distribution process will increase the time for communication significantly and affect the efficiency of the whole reconstruction process. This paper has proposed a tree data distribution method under MPI+CUDA and offered basic ideas and realization method of the algorithm. Experimental results have shown that tree distribution mode nearly doubles the speed than generally adopted plane distribution mode; its achievements provide certain reference for the development and processing of massive astronomical data.