I/O Performance Bounds for On-Array Processing on Mesh and Torus Networks

The aggregate beam-bandwidth capacity of digital phased arrays is typically constrained by data offload bottlenecks, preventing effective utilization and processing of most of the signal information received by a large sensor. On-array processing is an alternative architecture which aims to mitigate these bottlenecks by distributing data and processing loads across computing resources embedded directly within the array grid. This paper reviews optimal data-movement patterns on rectangular mesh and torus networks, and provides corresponding theoretical I/O-limited performance bounds for on-array processing under several different parallelization schemes.