Adaptive runtime systems for computational chemistry
American Chemical Society National Meeting & Exposition (ACS) 2014
Publication Type: Talk
Repository URL:
Summary
Computational modeling at the molecular level, carried out on supercomputers, promises to bring significant breakthroughs in areas related to chemistry, materials science, and biophysics. However, the upcoming era of ever more powerful machines and increasingly sophisticated algorithms brings with it new challenges, along with these promises. We have been developing an approach based on over-decomposition and adaptive runtime systems that we believe will be extremely useful in this era. Over-decomposition is the idea that the programmer decomposes the computation and data into a number of pieces much larger than the number of processors. The adaptive runtime system, which is in charge of assigning and reassigning these pieces to processors during execution, continuously instruments and introspect's the behavior of the application, and the machine. It then deploys several run-time including dynamic load balancing, to improve the performance of the running application. In addition to improving performance, this also leads to efficient composition of independently developed parallel modules. These ideas have been embodied in the parallel programming system called Charm++. In this talk, I will present our experience in developing highly scalable collaborative applications in biophysics (NAMD), and material modeling (OpenAtom), and describe how our methodology can be of use to the broader chemistry applications of the future. (248th ACS National Meeting and Exposition)
TextRef
People
Research Areas