An Artificial Neural Network based scheme for I/O FLASH AMR code optimization

Current modular Computational Fluid Dynamics (hereafter CFD ) codes often implement efficient parallel numerical schemes, resulting in efficient scalability scalable. The efficiency of the I/O phases however has received comparatively less attention. In this work we study the I/O performance of a highly structured community CFD-based multipurpose code FLASH (v.4.3). We have trained an Artificial Neural Network (hereafter ANN ) on a set of about 10000 I/O realistic runs, where the output files are sufficiently large and complex. We demonstrate that the ANN is quite successful in predicting the I/O efficiency in real time, given some input system status parameters.