Introduction Serial 1/0 on a parallel machine can have large time penalties for many reasons Larger datasets generated from parallel applications have a serial bottleneck if l/o is only done on one node Many MPP machines are built from large numbers of slower processors which increase the time penalty as the serial lo gets funneled through a single, slower processor Some parallel datasets are too large to be sent back to one node for file O Decomposing the computation phase while leaving the l /o channeled through one processor to one file can cause the time required for i o to be of the same order or exceed the time required for parallel computation There are also non-science applications in which input and output are the dominant processes and significant performance improvement can be obtained with parallel I /OIntroduction • Serial I/O on a parallel machine can have large time penalties for many reasons. – Larger datasets generated from parallel applications have a serial bottleneck if I/O is only done on one node – Many MPP machines are built from large numbers of slower processors, which increase the time penalty as the serial I/O gets funneled through a single, slower processor – Some parallel datasets are too large to be sent back to one node for file I/O • Decomposing the computation phase while leaving the I/O channeled through one processor to one file can cause the time required for I/O to be of the same order or exceed the time required for parallel computation. • There are also non-science applications in which input and output are the dominant processes and significant performance improvement can be obtained with parallel I/O