Satellite products computation with multiple GPU devices

Abstract

Over recent years, the number of earth observation satellites has increased dramatically. The satellite data gathering from various sources must be prepared and processed into satellite products (indexes), which takes computational time. Therefore, parallel computing techniques should be utilized in reducing time complexity. Graphic Processing Unit (GPU) devices are widely used to accelerate the computation process by massively parallel operations and, thus, are very suitable for this task. Although the index calculations were not complicated, the major drawback of the GPU process is data transfers between the host and the devices. Once data is transferred, it should be reused to calculate all related satellite indexes instead of beginning the transfer-compute cycle. In this paper, we investigate an efficacy way to produce satellite indexes and compare the computational times among many different approaches, i.e., Central Processing Unit (CPU) process executed based on NumPy and extended to multi-thread by Dask, single GPU device process performed with Numba, and multiple GPU device processes launched asynchronously. The experiments were carried out on three hardware environments, namely, the DGX workstation, the High-performance computing (HPC) High memory node, and the HPC-DGX node. The results revealed that the proposed GPU processes achieved more than ten times faster in overall. Furthermore, when compared with the CPU process, it found that its kernel computation could achieve more than 250 times faster.