# Communicator duplication (MPI)
When a C program calls a library, the library inherits all open file descriptors. If the library reads from `stdin` without knowing the application also reads from it, input gets silently consumed by the wrong side. The standard fix is to `dup()` the descriptor or close it before handing control over. MPI has the same problem at the message level. Every communicator has its own tag namespace: a message sent on one communicator cannot be matched by a receive on a different one, even between the same pair of ranks with the same tag. A library that uses `MPI_ANY_TAG` on `MPI_COMM_WORLD` can intercept messages the application sent to itself. `MPI_Comm_dup` creates a new communicator containing the same processes but with a fresh context, giving the library its own isolated namespace.

```c
MPI_Comm lib_comm;
MPI_Comm_dup(MPI_COMM_WORLD, &lib_comm);
library_init(lib_comm);   // all internal library traffic uses lib_comm
// ...
MPI_Comm_free(&lib_comm);
```

Passing `MPI_COMM_WORLD` directly into a third-party library is unsafe: the library's internal messages can be matched by wildcards in the application and vice versa, producing message mismatches that are difficult to diagnose. The convention is for any reusable MPI code to accept a communicator argument and duplicate it internally rather than operating on the passed communicator directly.