Dynamic Algorithm Selection in Parallel GAMESS Calculations

Abstract

Applications augmented with adaptive capabilities are becoming common in parallel computing environments which share resources such as main memory, network, or disk I/O. For large-scale scientific applications, dynamic adjustments to a computationally-intensive part may lead to a large pay-off in facilitating efficient execution of the entire application while aiming at avoiding resource contention. Application-specific knowledge, often best revealed during the run-time, is required to initiate and time these adjustments. In particular, general atomic and molecular electronic structure system (GAMESS) used for ab initio molecular quantum chemistry calculations has two different implementations of self-consistency field (SCF) methods, each of which targets either disk I/O or memory. This paper describes a mechanism enabling switching of algorithms during GAMESS run-time and shows the effect of the adaptations on the performance of GAMESS calculations as well as on a parallel GAMESS execution for different resource availability. The test results indicate that, in the presence of I/O resource contention, parallel GAMESS enhanced with adaptive mechanism may sustain the performance similar to that of full resource availability.