Exploiting the Cell Architecture on Roadrunner for Longtime Atomistic Dynamics
Authors:
Danny Perez
(Los Alamos National Laboratory)

Sriram Swaminarayan
(Los Alamos National Laboratory)

Arthur F. Voter
(Los Alamos National Laboratory)

Posters Session
Tuesday, 05:15PM  07:00PM

Room Rotunda Lobby

Abstract:
We show that Roadrunner, the new Cellbased massively parallel supercomputer being deployed at Los Alamos National Laboratory, is ideally suited to molecular dynamics simulations of small systems over very long timescales. Indeed, we demonstrate that we can achieve tenfold performance gain by limiting system sizes to a few thousand atoms such that all atoms and related bookkeeping information fit in the local store (LS) of the vector units (SPUs), as opposed to larger systems which require streaming atoms from main memory through the LS a few at a time. We show that this regime is ideally suited for the Parallel Replica Dynamics algorithm for the long time evolution of small systems. Thus, using the power of Roadrunner, the evolution of nanosystems over unprecedented timescales is accessible to direct molecular dynamics.