Improving parallel I/O autotuning with performance modeling