Parallel GLRT-Based SAR Tomographic Processing on Distributed Multi-GPU Platforms

This paper addresses the development of a multi-level parallel algorithm for detecting single scatterers in SAR tomography, targeted at multi-node multi-GPU distributed computational architectures. Taking into account the computational structure of the canonical processing scheme based on the Generalized Likelihood Ratio Test (GLRT) considered in this paper, an appropriate problem decomposition is adopted to devise and formulate an efficient parallel algorithm tailored for heterogeneous high-performance computing platforms, thereby enabling a hierarchical exploitation of different levels of parallelism, including both distributed and shared memory models. Experimental evaluations conducted on a multi-node, GPU-enabled HPC cluster, using a high-resolution SAR dataset, exhibit substantial acceleration and scalability. These results quantitatively confirm the effectiveness of the proposed multi-level parallel framework and position the developed prototype as a promising, scalable solution for large-scale SAR tomographic processing applications. This is particularly relevant in light of upcoming SAR missions, which are expected to generate unprecedented volumes of data, thus demanding scalable and high-performance processing solutions.