Molecular Dynamics Performance Evaluation with Modern Computer Architecture

Abstract

An important task of chemical biology is to discover the mechanism of recognition and binding between proteins. Despite the simplicity of the ligand-based model, fundamental mechanisms that regulate these interactions are poorly understood. An adequate equipment is mandatory to unravel this scientific challenge, not only through cost savings but also with high-quality results. With this in mind, we performed Molecular Dynamics simulations using the Gromacs package on two promising platforms: Cavium ThunderX2 ARM based cluster setup and shared-memory Intel based single-node machine. Aforementioned tests were also performed on common Intel based servers as a reference. Acquired results shown that shared-memory machine features the higest performance, although ARM and Intel clutsers are only slightly slower when more than four sockets are employed. During measurements, idle and job-execution consumptions were sampled in order to evaluate the energy required by a single simulation step. Results show that ARM and Intel servers are much less power-hungry with respect to shared-memory machine. The latter, on the other hand, features a decrement in power consumption when more resources are employed. Said unexpected behaviour is later discussed.