Transplantation and optimization of molecular dynamics simulation on MT-3000

Abstract

Molecular dynamics (MD) simulation is a research method to simulate the change of particle systems with the help of the powerful calculating power of computers. As the number of atoms increases and the duration of the simulation system extends, current mainstream software for MD simulations has become increasingly strained. This is largely due to a surge in computing time, attributed to the architecture characteristics of the central processing unit (CPU), which executes instructions serially and accesses shared memory randomly. These features have resulted in an enormous overhead that these software applications can no longer efficiently manage. Recently, National University of Defense Technology proposed the MT-3000 processor, a multi-zone heterogeneous architecture with a double-precision peak performance of 11.5 TFLOPS. In this paper, we implement the heterogeneous transplantation and optimization of the massively parallel program for MD simulation on Tianhe new-generation supercomputer based on MT-3000. By porting the key computing hotspot module to the acceleration zones of MT-3000, we have greatly improved the efficiency and performance of the simulation of billion-metal-atom cascade collision. The experimental results show that the heterogeneous MD program on a single node can achieve about 6 times faster than the original version executed in a general purpose zone of MT-3000. On the basis of 250 thousand cores, the strong-scalability parallel efficiency of heterogeneous MD program using 4 million cores can reach 89.89%, and the weak-scalability parallel efficiency can reach 91.98%. It can be predicted that the heterogeneous MD implementation has excellent scaling ability to simulate large-scale atomic cascade collisions on ten million of cores.