I/O distribution/reordering for more Contiguous Reads and Writes
S-01548
When sorting a file, for example, processes will request seemingly random list of traces during I/O. However, the whole application is requesting a contiguous chunk of the data file. By sorting and distributing the requests across the processes, it may be possible to recognize these as contiguous I/O calls, perform a contiguous I/O call, and then perform the more random communication between the processes (using e.g. MPI_Scatterv) instead of to/from the disk.
MPI / Lustre should be doing this under the hood, but there's really no guarantee that it is.
Old description:
Perform a contigious read and a contigous write of sorted traces using communication to transfer the traces and parameters.
CREATED ON - 18/11/2016