Performance of electronic structure calculations on BG/L and XT4 computers

The performance of the General Atomic and Molecular Electronic Structure System (GAMESS) on Cray XT4 and IBM Blue Gene (BG) supercomputers is presented. Efficient utilization of massively parallel computers enables the results of large self consistent field (SCF) and second order perturbation theory (MP2) calculations to be obtained in a short time. A number of performance limiting steps, including serial N³ steps in the orbital guess, SCF iterations, and property evaluations have been parallelized and optimized. Poor SCF cache performance has been addressed with moderate to significant improvements. Results of parallel N³ implementation on BG/L and XT4 systems, as well as on x86 clusters, are discussed. MP2 performance on BG/L and XT4 is compared and special considerations when running on BG/L are discussed. The improvements and discussions should be of interest to those considering using GAMESS or other electronic structure codes to study large systems on large supercomputers.