Ashok Srinivasan, Florida State University

Long-Time Molecular Dynamics Simulations through Parallelization of the Time Domain

Molecular Dynamics (MD) is an important atomistic simulation technique, with widespread use in computational chemistry, biology, and materials. An important limitation of MD is that the time step size is typically restricted to the order of femto (10^{-15}) seconds. Therefore a large number of iterations are required to reach realistic time scales. In fact, this is acknowledged as a major bottleneck in MD. While parallelization has been effective for dealing with the computational effort that arises in simulating large systems,  conventional parallelization is not effective in simulating small or  moderate sized physical systems to long time scales. We will present a new approach to parallelization that we recently introduced, where data from related simulations are used to parallelize a computation along the time domain. In this work, we use time parallelization to simulate a Carbon Nanotube, consisting of just around 1000 atoms, using MD, on up to 1000 processors with close to ideal speedups. In contrast, spatial decomposition is not effective on more than 2-3 processors. We also present results on Atomic Force Microscopy simulation of a protein, where this approach leads to an order of magnitude improvement in performance.

Ashok Srinivasan obtained a PhD in Computer Science at the University of California at Santa Barbara. He subsequently worked at NCSA, IIT Bombay, and UCSB, before joining Florida State University, where he is currently an Assistant Professor of Computer Science. His work has been supported by NSF and DoD, and computer time has been provided by NCSA, ORNL, NERSC, and IBM.