A general parallel solution to the integral transformation and second-order Moller-Plesset energy evaluation on distributed memory parallel machines

Abstract

We present here a parallel algorithm for four-index (integral) transformation and second-order Moller-Plesset (MP2) energy evaluation, primarily designed for multiple instruction multiple data (MIMD) machines. It is a general algorithm designed to work with equal efficiency on any inhomogeneous network and any architecture. This algorithm works with only a twofold redundancy in integral storage, whereas some previously reported strategies demand a fourfold redundancy. The parallel transformation and sorting algorithm has been implemented on a 128 node inhomogeneous ring network, 64 of the processors being about 30%-40% slower. The present parallel scheme is seen to perform excellently in integral transformation processes even in such an inhomogeneous environment due to dynamic load balancing strategies. It has been found that integral transformation along with MP2 energy evaluation takes typically 120-200 min for molecules with 80-90 atomic orbitals. However, time taken for such systems reduces to ~30-60 min when only requisite partial transformation is carried out for MP2 energy evaluation.