Hello,
I am working on a distributed array data structure built on top of the one-sided
communication primitives MPI_Get and MPI_Put in MPI-2. I use passive target
synchronization (MPI_Win_{lock,unlock}) in order to access remote entries in the array
without requiring the involvement of the target process. The MPI-2 standard allows
implementations to require that the memory used for this type of communication be
allocated using MPI_Alloc_mem. Hence, this is what I do, and it works well on my system.
I am, however, concerned that the amount of memory that can be allocated with
MPI_Alloc_mem could be limited on specific configurations, as stated in the following
article: <http://www.cs.berkeley.edu/~bonachea/upc/mpi2.html> . I have searched for
information regarding this issue, but not found anything. I would therefore like to ask
whether there are any such restrictions in OpenMPI, and in other popular implementations
(MPICH2, HP-MPI, etc)?
Kind regards,
Staffan Ronnas
