SC17 Denver, CO

P32: Exploring the Performance of Electron Correlation Method Implementations on Kove XPDs

Authors: Colleen Bertoni (Argonne National Laboratory), Brian Toonen (Argonne National Laboratory), William Allcock (Argonne National Laboratory), Spencer R. Pruitt (Worcester Polytechnic Institute), Mark S. Gordon (Iowa State University)

Abstract: In electron correlation methods in quantum chemistry, there are often high memory requirements which can reach terabytes for medium-sized molecular systems. For second-order perturbation theory (MP2), the two-electron integral arrays are the main memory bottleneck. Previously the two-electron integrals were recomputed, stored in distributed memory, or stored on disk. A way of storing the arrays which would remove the dependence on compute node memory and large latency associated with using disk is by using an external memory appliance, like Kove's XPD.

In this work. we modified a distributed memory implementation of MP2 to use XPDs instead of distributed memory. We evaluated the performance of our implementation against the distributed memory version for several molecular systems by considering scaling behavior with respect to compute processes and connections to XPDs. In the poster, we present an outline of the MP2 implementation using XPDs and the scaling results.

Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf

Poster Index