MPIPut

The term MPIPut refers to the MPI_Put routine that is part of the MPI one‑sided communication, or Remote Memory Access (RMA) model. MPI_Put allows a process, called the origin, to write data directly into the memory window of another process, called the target, without any explicit cooperation from the target at the time the data is transmitted. The routine completes a data transfer by performing a local copy on the origin host and then a remote copy to the target window.

The MPI_Put call does not alter program flow or require a matching receive operation on the target, but the operation is only visible to the target after the window is synchronized. Before performing a Put, the origin must obtain a lock on the target window using MPI_Win_lock or MPI_Win_lock_all with the MPI_LOCK_SHARED or MPI_LOCK_EXCLUSIVE mode. After all pending puts have been issued, the origin calls MPI_Win_flush, MPI_Win_unlock, or MPI_Win_unlock_all to ensure the data becomes visible. When combined with the exposure epoch, synchronization guarantees that the target will see the data in a consistent state, preventing race conditions.

The syntax for MPI_Put is: MPI_Put(void *origin_addr, int origin_count, MPI_Datatype origin_datatype, int target_rank, MPI_Aint target_disp, int target_count, MPI_Datatype target_datatype, MPI_Win win). Here origin_addr points to the source buffer, origin_count and origin_datatype describe the data to be sent, target_rank identifies the destination process, target_disp specifies the displacement within the target window, target_count and target_datatype describe the target buffer, and win is the window object.

MPI_Put is often used in scientific codes that implement global arrays on distributed memory, enabling efficient direct updates of elements across processes. It delivers low communication overhead when combined with advanced window types like MPI_Win_create_dynamic. The routine can be used in blocking or non‑blocking fashion; non‑blocking calls are indicated by the additional MPI_Put function. It is a fundamental building block in modern parallel programming with MPI, offering fine‑grained control over data placement and transfer.