基于Dask并行加速的射电干涉成像网格化方法实现

快速傅里叶变换(Fast Fourier Transform,FFT)比傅里叶变换有更好的算法性能,是射电干涉成像的基础算法,但因为天线阵列的不规则采样,需使用网格化算法将可见度数据重采样到规则的网格上才能应用。基于卷积的网格化计算具有密集型和迭代型的特点,特别是处理海量可见度数据的情况下,高性能的网格化计算对整个成像过程加速尤为重要。为了缓解数据处理的压力,在现有处理整块数据和支持多核计算的算法基础上,拓展应用Dask并行计算框架,不仅将数据分块并分配到多线程上,提高数值计算效率,而且动态的分布式任务调度策略优化了网格化的实时处理。实验结果表明,多核中央处理器利用率显著提高,即使增加数据量,也能进一步提高网格化算法的性能。分布式任务调度能够将单(多)测量集的网格化弹性缩放到单(多)机系统,充分发挥集群的规模化优势。

A Distributed Gridding Implementation Method for Radio Interferometric Visibilities Based on Dask

Fast Fourier Transform(FFT)has better performance than Discrete Fourier Transform,which is the fundamental imaging algorithm of radio interferometry.However,because of the irregular sampling of antenna array,it is necessary to use gridding algorithms to resample visibilities to regular grids,so that FFT can be applied.The convolutional gridding in radio interferometric imaging is characterized by intensive and iterative computations.Especially in the case of massive visibility data processing,high-performance gridding computing is particularly important to accelerate the whole imaging process.In order to alleviate the pressure of data processing,the Dask parallel computing framework is extended and applied on the existing gridding algorithm which supports multi-core parallelism but processes whole blocks of data.Not only can the data be partitioned and distributed to multiple threads to improve the efficiency of numerical computation,but also the dynamic distributed task scheduling strategy can optimize the real-time workflow of gridding.The experimental results show that the multi-core utilization rate is significantly improved and the performance of gridding algorithm can be further enhanced even if the volume of visibility is increased.Distributed task scheduling can flexibly scale the gridding task of(single)multi-measurement set to(single)multi-machine system,which gives full play to the scale advantage of clustering.