On-the-fly clustering for exascale molecular dynamics simulations

Computational resources have experienced exponential growth in the last decades enabling the simulation of complex physical problems at the cost of a massive increase in data storage. This is especially true for N-body simulations now reaching billions or trillions particles in certain cases. To overcome the drawbacks of data storage on disk for post-processing purposes, on-the-fly analysis has gained momentum but still represents a challenge in both its implementation and efficiency without impacting the simulation engine performances. This work provides a new in-situ procedure for features detection in massive N-body simulations, leveraging state-of-the-art techniques from various fields. Based on a discrete-to-continuum paradigm shift, particles and their respective physical quantities are projected onto a 3D regular grid before applying image analysis algorithms to group voxels based on specific user-defined criteria. A significant extension to the hybrid parallelism of connected component analysis within the image processing community is also introduced in the present study. Traditionally operating in shared memory parallelism, this extension now incorporates both distributed and shared memory approaches. The implementation is carried out within the exaStamp classical Molecular Dynamics code, a variant of the open-source exaNBody platform [39]. This adaptation allows for the on-the-fly analysis of multi-billion atoms samples with at most a 1.3% overhead. In addition, the entire framework is benchmarked up to 32768 cores. The applicability of the present approach is demonstrated on the case of a spall fracture in a tantalum sample as well as high velocity impact of a tin droplets on a rigid surface.