On-the-fly clustering for exascale molecular dynamics simulations.

Connected components analysis can be categorized into two primary branches: those developed by the image processing community and those from the graph processing community.

The image processing community primarily utilizes shared memory algorithms that operate with a finite number of passes, which limits their scalability to a single node. To effectively employ these algorithms, it is necessary to integrate them into a structured memory system.

In contrast, the graph processing community has developed algorithms that are fully parallelized and typically involve many iterative processes. However, these algorithms require adaptation for application to 3D images.

The emphasis on images arises from the need to work with atomic data, which consists of moving atoms. However, our focus will shift from atomic data to projecting these quantities onto a regular grid. This involves subdividing the spatial domain into cells, allowing us to define subcells or box cells for our analysis. Each atomic quantity will be projected onto these boxes, resulting in a 3D voxelized image. Our initial approach will be to implement the methodologies from the image processing community.