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

[Jeff Hammond]

I'd like to be able to use MPI_Datatypes to send e.g. 4 GiB messages.  Right now, even 2 GiB fails.

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

Thanks,

Jeff

[jhammond at protos mpi64]$ ./test_type_contig.x 15
count = 32768 
ocount = 32768 
[jhammond at protos mpi64]$ ./test_type_contig.x 20
count = 1048576 
ocount = 1048576 
[jhammond at protos mpi64]$ ./test_type_contig.x 25
count = 33554432 
ocount = 33554432 
[jhammond at protos mpi64]$ ./test_type_contig.x 30
count = 1073741824 
ocount = 1073741824 
[jhammond at protos mpi64]$ ./test_type_contig.x 31
count = 2147483648 
Assertion failed in file ../src/mpid/common/datatype/dataloop/segment_packunpack.c at line 133: FALSE
memcpy argument memory ranges overlap, dst_=0x254b310 src_=0x7f6ad81f3010 len_=-2147483648

internal ABORT - process 0
[jhammond at protos mpi64]$ ./test_type_contig.x 32
count = 4294967296 
ocount = 0 

-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond
ALCF docs: http://www.alcf.anl.gov/user-guides