Augmenting BEACON Processing Elements with Sorting Support

Determine whether the BEACON processing elements can be augmented to support sorting algorithms, such as Shearsort, for the sorting operations required during exact k-nearest-neighbor processing.

Background

BEACON accelerates exact two-dimensional k-nearest-neighbor construction by computing Euclidean distances, assigning distances to buckets, aggregating bucket counts, and sorting the distances in a selected bucket to identify the final nearest neighbors. In the described design, this final sorting step is performed at the global processing element using a multi-step linear scan.

The paper leaves unresolved whether sorting should instead be implemented directly within the processing-element array. It specifically identifies Shearsort as a potentially suitable algorithm because the accelerator uses a two-dimensional array of processing elements, but does not establish whether the processing elements can be augmented to support it efficiently. Such support could reduce the cost of the final bucket-sorting stage in k-nearest-neighbor processing.