Scaling Radio Astronomy Signal Correlation on Heterogeneous Supercomputers Using Various Data Distribution Methodologies

TL;DR

This paper evaluates different data distribution strategies for GPU-based signal correlation in next-generation radio telescopes, focusing on scalability and throughput using the Fornax supercomputer.

Contribution

It systematically compares multiple data distribution models for GPU clusters, identifying optimal approaches for large-scale radio astronomy signal correlation.

Findings

Optimal data distribution models vary with problem size.

GPU cluster implementations achieve high scalability and throughput.

Systematic testing guides efficient design for next-generation telescopes.

Abstract

Next generation radio telescopes will require orders of magnitude more computing power to provide a view of the universe with greater sensitivity. In the initial stages of the signal processing flow of a radio telescope, signal correlation is one of the largest challenges in terms of handling huge data throughput and intensive computations. We implemented a GPU cluster based software correlator with various data distribution models and give a systematic comparison based on testing results obtained using the Fornax supercomputer. By analyzing the scalability and throughput of each model, optimal approaches are identified across a wide range of problem sizes, covering the scale of next generation telescopes.