Fault Tolerant Molecular-Continuum Flow Simulation

Molecular-continuum simulations couple molecular dynamics (MD) and computational fluid dynamics (CFD) simulations in a domain decomposition sense to assess fluid flow, e.g., in process engineering applications, at the nanoscale. Running these simulations on extreme-scale supercomputers, an issue consists in single compute cores or nodes failing due to hardware- or software-sided errors. This imposes a challenge to robustness of numerical simulations and, as such, also to molecular-continuum systems. We introduce a fault tolerance method in our macro-micro-coupling tool (MaMiCo) that has been developed in the past as molecular-continuum simulation software solution. With MaMiCo leveraging ensemble simulations to cope with statistical errors in the MD solutions, we extended the ensemble approach to recognize failing MPI processes and react to these failures. Once a failure is encountered, the affected MD simulations are removed from these MPI processes and relaunched on well-operating MPI process groups. We detail our approach and report scalability results for our approach, achieved on the supercomputer HAWK at HLRS.