Signal and Image processing

Exploring Algorithmic and Distributed HPC Strategies for Radio Astronomy Using SimSDP

Published on - Astronomy and Computing

Authors: Ophélie Renaud, Sunrise Wang, Nicolas Gac

Modern radio telescopes, such as the SKAO telescopes, generate massive data streams that must be processed efficiently on large-scale HPC systems. Designing scalable imaging pipelines and estimating the computing resources required for future facilities remains challenging due to the complexity of radio-interferometric algorithms and the diversity of possible hardware architectures. In this work, we present an extended version of SimSDP, a rapid prototyping and simulation tool for radio astronomy pipelines. The proposed prototyping tool enables both algorithmic exploration and HPC resource estimation by combining dataflow-based modeling, parametric execution time models, and distributed system simulation. We extend SimSDP to support multi-node and GPU architectures with precise hardware calibration. New distribution and parallelization strategies are introduced, including decentralized processing schemes and models accounting for non-uniform visibility distributions across compute nodes. Parametric timing models, depending on key algorithmic parameters, allow simulations to scale toward SKAO-level workloads. This tool is validated through extensive empirical experiments involving 4320 pipeline configurations executed as distributed prototypes on CPU and GPU platforms. Simulation results are compared with measured execution times, demonstrating that SimSDP can reliably predict performance trends and support the design and sizing of future radio astronomy HPC systems.