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

You begin with your unique voxel idea and proceed to label each voxel. Start with the first voxel, labeled as 0, and assign it the label 1. Continue this process for the second, third, fourth, fifth, and sixth voxels. The sixth voxel is isolated and currently unconnected, prompting the definition of a new connected component or level.

Next, move to the second line. The sixth voxel is now connected to a voxel from the first level, allowing you to assign it the same label. This process facilitates the creation of global graphs. However, at this stage, not all voxels are interconnected, necessitating a final pass to consolidate the information and unify all levels to the minimum label assigned.

The objective is to achieve DOMAIN labeling, which involves the interaction between threads and MPI. The DOMAIN labeling between threads is relatively straightforward. We will now focus on the more complex MPI DOMAIN labeling.

To implement this, we will utilize a master MPI to facilitate the stitching process. Each MPI will send messages to the master MPI, enabling it to perform the global connection. For illustration, consider a scenario with one MPI and one NKI, where the local graphs are constructed by the NPI, as shown on the right.