Performance of communication patterns for extreme-scale analysis and visualization

Efficient data movement is essential for extreme-scale parallel visualization and analysis algorithms. In this research, we benchmark and optimize the performance of collective and point-to-point communication patterns for data-parallel visualization of scalar and vector data. Two such communication patterns are global reduction and local nearest-neighbor communication. We implement scalable algorithms at tens of thousands of processes, in some cases to the full scale of leadership computing facilities, and benchmark performance using large-scale scientific data.