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.
# Scatter and gather (MPI)
The basic pattern for distributing work is: rank 0 holds the full dataset, divides it into N equal chunks, and sends one to each process. Without collectives, that is a loop of sends:

```c
if (rank == 0)
    for (int i = 0; i < size; i++)
        MPI_Send(&data[i * chunk], chunk, MPI_INT, i, 0, MPI_COMM_WORLD);
else
    MPI_Recv(local, chunk, MPI_INT, 0, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
```

This serialises all N sends on rank 0. `MPI_Scatter` expresses the same pattern as a single collective call that all processes participate in simultaneously, dividing a root's buffer into equal-sized chunks and delivering one to each process including the root itself. `MPI_Gather` is the inverse: each process sends a chunk back and root assembles the result.

```c
int send_buf[size];  // root fills this
int recv_val;
MPI_Scatter(send_buf, 1, MPI_INT, &recv_val, 1, MPI_INT, 0, MPI_COMM_WORLD);
```

`MPI_Scatterv` and `MPI_Gatherv` are the variable-length variants, where each process sends or receives a different number of elements. They take a counts array and a displacements array, so chunks do not need to be equal in size or contiguous in the root's buffer. This is the right tool for distributing irregular domain decompositions where the workload per process is not uniform.