[mpich-discuss] MPICH datatype memory utilization and low memory system

[Rob Latham]

We've encouraged our HDF5 friends to use MPI datatypes to describe their 
I/O with the promise that "we'll fix it" if they find problems in the 
MPI implementation.

the MOAB folks use HDF5 to store and read their unstructured mesh data, 
and so HDF5 is using MPI_Type_hindexed to describe these approximately 
16 million points.

ON blue gene... well you maybe can guess where I'm going with this. 
MPICH will use quite a bit of memory to describe 16 million points: 
about 200 MiB (4 bytes for each index, 8 bytes for each offset).

there isn't 200 MiB of extra memory on some configurations of Blue Gene.

What options does one have when trying to efficiently describe 
unstructured data, though?

==rob

-- 
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA