RadioLensfit: an HPC Tool for Accurate Galaxy Shape Measurement with SKA

TL;DR

RadioLensfit is an HPC-optimized, open-source tool that efficiently measures galaxy shapes in radio weak lensing studies using SKA data, leveraging Fourier domain fitting and parallel computing for large datasets.

Contribution

It introduces RadioLensfit, a novel Fourier domain fitting method with hybrid MPI+OpenMP parallelization for fast, accurate galaxy shape measurement in radio weak lensing.

Findings

Achieves measurement accuracy comparable to previous methods.

Significantly reduces computational time for large datasets.

Demonstrates scalability on SKA-MID simulated data.

Abstract

The new generation radio telescopes, such as the Square Kilometre Array (SKA), are expected to reach sufficient sensitivity and resolution to provide large number densities of resolved faint sources, and therefore to open weak gravitational lensing observations to the radio band. In this paper we present RadioLensfit, an open-source tool for an efficient and fast galaxy shape measurement for radio weak lensing shear. It performs a single source model fitting in the Fourier domain, after isolating the source visibilities with a sky model and a faceting technique. This approach makes real sized radio datasets accessible to an analysis in this domain, where data is not yet affected by the systematics introduced by the non-linear imaging process. We detail the implementation of the code and discuss limitations of the source extraction algorithm. We describe the hybrid parallelization MPI+OpenMP of the code, implemented to exploit multi-node HPC infrastructures for accelerating the computation and dealing with very large datasets that possibly cannot entirely be stored in the memory of a single processor. Finally, we present performance results both in terms of measurement accuracy and code scalability on SKA-MID simulated datasets. In particular, we compare shape measurements of 1000 sources at the expected source density in SKA Phase 1 with the ones obtained from the same dataset in a previous work by a joint fitting of the raw visibility data, and show that results are comparable while the computational time is highly reduced.