Ivan's wiki

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
# One-sided communication (MPI)
`mmap` with `MAP_SHARED` lets two processes on the same machine map the same physical memory. One process writes through its pointer and the other immediately sees the update — no explicit send or receive, no coordination required from the reader. One-sided MPI communication (also called **RMA**, remote memory access) extends this idea across nodes. A process exposes a memory region as a **window** with `MPI_Win_create`, and any other process in the communicator can write into it with `MPI_Put` or read from it with `MPI_Get`, without the target process calling any MPI function.

```c
int val = (rank == 0) ? 42 : 0;
MPI_Win win;
MPI_Win_create(&val, sizeof(int), sizeof(int), MPI_INFO_NULL, MPI_COMM_WORLD, &win);
MPI_Win_fence(0, win);
if (rank == 1)
    MPI_Get(&val, 1, MPI_INT, 0, 0, 1, MPI_INT, win);  // rank 1 reads rank 0's val
MPI_Win_fence(0, win);   // val on rank 1 is now 42
MPI_Win_free(&win);
```

`MPI_Win_fence` acts as a collective barrier that opens and closes access epochs. `MPI_Accumulate` is an atomic RMA update that applies an MPI operator rather than simply overwriting. Finer-grained passive-target synchronisation is possible with `MPI_Win_lock` and `MPI_Win_unlock`, where the target process does not participate at all.