MPIDatatype

MPIDatatype is an internal representation of MPI datatypes used by MPI implementations, such as MPICH and Open MPI, to describe the layout of data in communications. It provides the concrete object behind the MPI_Datatype handle and encompasses both built-in types and user-defined derived types.

The main purpose of MPIDatatype is to support the communication subsystem in determining how many elements to send, how much memory to read, where to start packing, and how datatype information should be interpreted across processes. It captures the memory layout requirements of a datatype, supporting heterogeneous architectures and enabling correct and efficient data transfer.

While the exact fields are implementation-specific, a MPIDatatype generally stores information such as the datatype’s size, extent, and true extent, and whether the type is contiguous. For derived types, it maintains references to constituent types and describes the layout (for example, strides, block lengths, and displacement arrays for vector and indexed types, or field entries for struct types). This internal description allows the runtime to perform packing, unpacking, and type matching during communication.

Lifecycle and usage follow the MPI datatype workflow. Derived MPIDatatypes are created by MPI_Type_* constructors and finalized with MPI_Type_commit; built-in types have a corresponding internal representation. MPIDatatype objects are typically reference-counted and are freed when MPI_Type_free releases the last reference. Implementations may cache or reuse internal representations to optimize performance.

Notes and scope: MPIDatatype is an abstraction whose exact layout is implementation-dependent. Nevertheless, the concept across MPI implementations is consistent: map a datatype handle to an internal description that supports correct data transfer, type matching, and efficient packing/unpacking across processes.

See also: MPI datatype, MPI_Type_commit, MPICH, Open MPI.