Molecular Dynamics at its Limits: Scalable Algorithms, HPC Software & Multiscale Methods for Nanoflow Simulation

Molecular dynamics (MD) has evolved as a powerful tool in engineering. However, MD considerations often require the simulation of large molecular systems with millions to billions of atoms, or even more; examples comprise investigations of molecular droplet coalescence, bubble formation, or flows in nano-sized devices such as nanofilters.
Numerical simulation of these scenarios necessitates efficient algorithms, optimal implementations thereof and the use of high-performance computing (HPC) resources, with the latter becoming more and more complex: heterogeneous compute nodes and higher levels of parallelism due to increasing node and core counts are only two out of many challenges that have to be faced.

In my talk, I cross the worlds of algorithms, HPC and software development by discussing simulation methods that are relevant for the aforementioned application examples.
I present algorithms for short-range MD that scale well on shared-memory platforms such as multi- and manycore systems. These methods, in combination with optimal MPI communication and single-core implementations, allow for the first time the simulation of 20 trillion atoms, running at 1.3 PFLOPS on 172 000 cores.
I further detail the limits of MD and provide a remedy for selected flow problems in terms of molecular-continuum flow simulations. Molecular-continuum systems leverage the multiscale idea and combine computational fluid dynamics with MD in a smart way to enable the investigation of flows at large (spatial and temporal) scales. I present a transient algorithm, enhanced by multi-instance MD sampling, that enables large scale molecular-continuum flow simulation on up to 65000 compute cores.