Distributed Asynchronous Union-Find for Scalable Feature Tracking.

Feature tracking and the visualizations of the resulting trajectories make it possible to derive insights from scientific data and thus reduce the amount of data to be stored. However, existing serial methods are not scalable enough to handle fast increasing data size. In this paper, we tackle the problem of distributed parallelism for feature tracking and visualization by introducing a scalable, asynchronous union-find algorithm. We show that asynchronous communication can improve the scalability of distributed union-find operation in comparison to synchronous communication, as seen in existing methods. In the proposed feature tracking pipeline, we first construct and partition a high-dimensional mesh that incorporates both space and time. Then, the trajectories of features are built distributively across parallel processes, and trajectory pieces are merged asynchronously by using our distributed union-find implementation. Results demonstrate the scalability of tracking critical points on exploding wire experimental data and tracking super level sets on BOUT++ fusion plasma simulations.