Fast Distribtued Selection with Graphics Processing Units
Ruizhe Fu † Jeffrey D. Blanchard *, and Tristan Knoth ‡
IEEE Open Journal of the Computer Society, 2024
We introduce two completely parallel algorithms for exactly solving the distributed multiple order statistic problem. The main algorithm, DIBMS, keeps communication costs proportional to the number of order statistics with minimal communication of actual data. Moreover, the work done at the remote nodes is fully parallelized on graphics processing units.