Data Multiplexing in Radio Interferometric Calibration

TL;DR

This paper introduces a multiplexing calibration scheme for radio interferometry data that reduces computational resource requirements by enabling simultaneous calibration across all frequencies with fewer compute agents.

Contribution

It proposes an ADMM-based multiplexing method that allows calibration of full datasets using fewer compute agents, improving efficiency in radio astronomy data processing.

Findings

Simulation confirms feasibility of the multiplexing scheme

Enables calibration with limited computational resources

Improves processing efficiency for large radio datasets

Abstract

New and upcoming radio interferometers will produce unprecedented amounts of data that demand extremely powerful computers for processing. This is a limiting factor due to the large computational power and energy costs involved. Such limitations restrict several key data processing steps in radio interferometry. One such step is calibration where systematic errors in the data are determined and corrected. Accurate calibration is an essential component in reaching many scientific goals in radio astronomy and the use of consensus optimization that exploits the continuity of systematic errors across frequency significantly improves calibration accuracy. In order to reach full consensus, data at all frequencies need to be calibrated simultaneously. In the SKA regime, this can become intractable if the available compute agents do not have the resources to process data from all frequency channels simultaneously. In this paper, we propose a multiplexing scheme that is based on the alternating direction method of multipliers (ADMM) with cyclic updates. With this scheme, it is possible to simultaneously calibrate the full dataset using far fewer compute agents than the number of frequencies at which data are available. We give simulation results to show the feasibility of the proposed multiplexing scheme in simultaneously calibrating a full dataset when a limited number of compute agents are available.