Atomistic simulation of low-dimensional nanostructures toward extreme-scale supercomputing

Large-scale atomistic simulation of low-dimensional silicon nanostructures has been implemented on a heterogeneous supercomputer equipped with a large number of GPU-like accelerators (GLA). In the simulation, an innovative parallel algorithm was developed for the combined utilization of the dynamic neighbor and static neighbor list algorithms aiming at the different regions of the nanostructures. Furthermore, some optimization techniques were performed for the computationally intensive many-body force evaluation between atoms, such as SIMD vectorization, manual loop unrolling, pre-calculation of memory addresses and reordering of data structure etc. Finally, the simulation achieved the excellent weak and strong scalabilities in the parallel implementation, where up to 805.3 billion silicon atoms were simulated. This development suggests an exciting future of predicting the thermodynamic properties of low-dimensional nanostructures.