[mpich-discuss] halo update performance comparison

[Kokron, Daniel S. (GSFC-610.1)[Computer Sciences Corporation]]

All,

I have been interested in MPI datatypes since reading about them and playing with the benchmark found at http://unixer.de/research/datatypes/ddtbench/

I am working with a weather model that spends considerable time doing halo (aka ghost cell) updates.  The existing code uses irecv/isend.  Since most of the time spent doing halo updates is actually spent packing and unpacking the MPI buffers, I was hoping the use of datatypes would improve performance.  Unfortunately, the datatype version is actually quite a bit slower than the standard non-blocking pt2pt code.  I found this to be that case for both mpich-3.1rc1 and a recent version of a vendor MPI implementation (not MPICH based).  I have not run this comparison using OpenMPI.

Is data type performance an active area of development?  Is there room for improvement?

Daniel Kokron
NASA Ames (ARC-TN)
SciCon group
301-286-3959