get_address fails on arm

[Thomas Spura]

It seems that there is an failure on arm with openmpi:

FAIL: testGetAddress (test_datatype.TestGetAddress)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test/test_datatype.py", line 360, in testGetAddress
    self.assertEqual(addr, bufptr)
AssertionError: -2128074816 != 2166892480L
------------------------------------------------

It seems some cython type might be wrong, but actually this could also be a MPI failure and not mpi4py's...

The full log can be found here.

[Lisandro Dalcin]

This is a signed vs unsigned integer issue. MPI_Get_address() returns addresses as MPI_Int (signed integral type), while Python is returning addresses as unsigned integer (but this is not consistent across Python versions, however I should check the status of this for Py26+). See yourself:

>>> from struct import *
>>> unpack('i', pack('I', 2166892480))
(-2128074816,)

In other words, the test case is broken and should be fixed.

[Thomas Spura]

It seems MPI_Int has a different size on arm than on the other arches. Does this imply, that it is not possible to mix different arches in a multi node MPI program? (I'm just asking being curious, not if such a call would make sense ;)) I always expected to have MPI_Int of the same size accross architectures.

[Lisandro Dalcin]

Sorry, my previous comment is misleading. I was talking about MPI_Aint (not the A in the name).

MPI_Aint should be equivalent to intptr_t. IOW, MPI_Aint is a signed integral type large enough to represent any address, e.g., in i386 MPI_Aint should be plain C int.

About muti-arch MPI runs (heterogeneous environments in the parlance of the MPI standard), that is perfectly possible, though the MPI implementation has to provide support for it.

[Lisandro Dalcin]

Test fixed in commit 4a4df39 and 259f09b