[Feature Request] Parallel HDF5 Read-Proc0-And-Bcast Improvements

[jhendersonHDF]

A few ideas were noted for improvements that could be made to parallel HDF5's solution for optimizing the case where "small" amounts of data get read by all MPI processes:

• It should be possible to optimize the algorithm to broadcast both the dataset elements and the # of bytes read in one operation instead of two.
• It could be useful to add a DXPL property to let the application know that this optimization was invoked.
• It should be investigated whether this optimization can be generalized to more cases besides just H5S_ALL selections on contiguous datasets of size <= 2GiB. This would likely involved detecting regions of selections that are common to all MPI processes and optimizing the I/O for those regions.