Efficient input and output for scientific simulations

Large simulations which run for hundreds of hours on paralle l computers often periodically generate snapshots of states, wh ich are later post-processed to visualize the simulated physical p henomenon. For many applications, fast I/O during post-processing, wh ich is dependent on an efficient organization of data on disk, is as i mportant as minimizing computation-time I/O. In this paper we pr opose optimizations to support efficient parallel I/O for scienti fic simulations and subsequent visualizations. We present an orderin g mechanism to linearize data on disk, a performance model to help t o choose a proper stripe unit size, and a scheduling algorithm to inimize communication contention. Our experiments on an IBM S P show that the combination of these strategies provides a 2025% performance boost.