Parallel short range molecular dynamics simulations on computer clusters: Performance evaluation and modeling

This paper describes the performance of a portable molecular dynamics code running on an eight-node PC cluster. The molecular dynamics code is based on the atom decomposition method for distributing the computation load among the processors and the MPI protocol for managing communications among processors. We discuss the changes made to the serial code with an effort to maintain its readability. We examined the program performance for system sizes of order 102 to 104 atoms and number of processors varying from 1 to 8, by measuring the total execution time and the corresponding speedup, as well as the communication time for data exchange and the time for the calculation of interatomic forces. Using simple communication and computation load considerations, we propose models in order to explain the observed behaviour and predict the optimal usage of the cluster. It turns out that using few parameters that can be easily measured one can predict quite accurately the optimal usage of small clusters running short range molecular dynamics programs.