This ticket proposes a new function that allows the user to specify an info key for MPI datatypes.
The motivation for this is:
Most other MPI objects (e.g Comm, Win, File) permit info keys.
Implementations that attempt to optimize datatypes will be able to use this, e.g. an info key can hint at how many times the datatypes will be reused, which will motivate more optimization and vice versa. There is little reason to optimize a datatype that will only be used once, for example.
The new functions proposed are
MPI_TYPE_COMMIT_WITH_INFO(datatype, info)
INOUT datatype - datatype that is committed (handle)
IN info - associated info key
MPI_TYPE_GET_INFO(datatype, info)
IN datatype - datatype (handle)
INOUT info - associated info key
The reason that MPI_TYPE_DUP_WITH_INFO is not desirable is that the overhead of doing this is non-negligible as compared to MPI_COMM_DUP. Furthermore, one of the motivations is to affect how the commit process occurs, so this information should be present at this time.
This ticket proposes a new function that allows the user to specify an info key for MPI datatypes.
The motivation for this is:
The new functions proposed are
The reason that
MPI_TYPE_DUP_WITH_INFO
is not desirable is that the overhead of doing this is non-negligible as compared toMPI_COMM_DUP
. Furthermore, one of the motivations is to affect how the commit process occurs, so this information should be present at this time.