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.
# Parallel I/O (MPI)
POSIX `pwrite(fd, buf, n, offset)` lets multiple processes write to different offsets in the same file without interfering: the offset is passed explicitly rather than shared, so there is no race on the file position. The straightforward approach in a parallel program is for each process to compute its own offset and call `pwrite` directly.

MPI-IO builds on the same idea but adds a collective interface: each process opens the same file with `MPI_File_open` and writes its portion with an explicit offset. The advantage of the collective call `MPI_File_write_at_all` is that the runtime can merge small contributions from all processes into fewer, larger I/O operations aligned to filesystem stripe boundaries.

```c
MPI_File fh;
MPI_File_open(MPI_COMM_WORLD, "output.bin",
              MPI_MODE_CREATE | MPI_MODE_WRONLY, MPI_INFO_NULL, &fh);

MPI_Offset offset = (MPI_Offset)rank * N * sizeof(double);
MPI_File_write_at_all(fh, offset, local_data, N, MPI_DOUBLE, MPI_STATUS_IGNORE);

MPI_File_close(&fh);
```

File views (`MPI_File_set_view`) let each process define its portion of the file using a derived datatype so that subsequent read/write calls use a logical offset within the view rather than raw byte offsets. This is the standard approach for writing distributed arrays: each process defines a view covering its slice of the global array, then calls `MPI_File_write` as if writing a contiguous local buffer.