[mpich-devel] when is MPICH going to be count-safe?

[Rob Latham]

On Sun, Oct 13, 2013 at 09:13:31PM -0500, Jeff Hammond wrote:
> I'd like to be able to use MPI_Datatypes to send e.g. 4 GiB
> messages.  Right now, even 2 GiB fails.

Right on, man.  My "large count large type" branch has been stuck in
review purgatory.  Think you are you qualified to review
mpich-review/large_count_large_typesize_combo ?

> Code is here https://github.com/jeffhammond/HPCGloves/tree/master/examples/mpi64.  Test machine is in MCS.

er, maybe I should say "qualified to review, once I pass your tests":

% ./test_type_contig.x 15
count = 32768 
ocount = 32768 
% ./test_type_contig.x 25
count = 33554432 
ocount = 33554432 
% ./test_type_contig.x 30
count = 1073741824 
ocount = 1073741824 
% ./test_type_contig.x 31
count = 2147483648 
ocount = -2147483648 

Heh.  silent error vs violent explosion.  I'll let you consider which
mode is preferable while I go fix my branch.

==rob

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