Radio-Astronomical Imaging on Accelerators

Imaging is considered the most compute-intensive and therefore most challenging part of a radio-astronomical data-processing pipeline. To reach the high dynamic ranges imposed by the high sensitivity and large field of view of the new generation of radio telescopes such as the Square Kilometre Array (SKA), we need to be able to correct for direction-independent effects (DIEs) such as the curvature of the earth as well as for direction-dependent time-varying effects (DDEs) such as those caused by the ionosphere during imaging. The novel Image-Domain gridding (IDG) algorithm was designed to avoid the performance bottlenecks of traditional imaging algorithms. We implement, optimize, and analyze the performance and energy efficiency of IDG on a variety of hardware platforms to find which platform is the best for IDG. We analyze traditional CPUs, as well as several accelerators architectures. IDG alleviates the limitations of traditional imaging algorithms while it enables the advantages of GPU acceleration: better performance at lower power consumption. The hardware-software co-design has resulted in a highly efficient imager. This makes IDG on GPUs an ideal candidate for meeting the computational and energy efficiency constraints of the SKA. IDG has been integrated with a widely-used astronomical imager (WSClean) and is now being used in production by a variety of different radio observatories such as LOFAR and the MWA. It is not only faster and more energy-efficient than its competitors, but it also produces better quality images