Symmetrical Data Sieving for Noncontiguous I/O Accesses in Molecular Dynamics Simulations

摘要

This article analyzes the impact of different solutions of I/O performance problem on Molecular Dynamics (MD) simulation. MD accesses data in a non-contiguous way, with unpredictable patterns that change as the simulation progresses. Such application drastically affect the performance of Parallel File Systems. We show that there is a strong relation between the distribution of the atoms on the simulation space and MD I/O performance. We propose a variant of data sieving approach that uses the symmetry of the MD's access pattern to minimize the number of I/O operations, and behaves considerably better than other tested solutions.