8.8. Naming Objects

There are many occasions on which it would be useful to allow a user to associate a printable identifier with an MPI communicator, window, or datatype, for instance error reporting, debugging, and profiling. The names attached to opaque objects do not propagate when the object is duplicated or copied by MPI routines. For communicators this can be achieved using the following two functions.

MPI_COMM_SET_NAME allows a user to associate a name string with a communicator. The character string that is passed to MPI_COMM_SET_NAME will be saved inside the MPI library (so it can be freed by the caller immediately after the call, or allocated on the stack). Leading spaces in name are significant but trailing ones are not.

MPI_COMM_SET_NAME is a local (noncollective) operation, which only affects the name of the communicator as seen in the MPI process that made the MPI_COMM_SET_NAME call. There is no requirement that the same (or any) name be assigned to a communicator in every MPI process where it exists.

Advice to users.

Since MPI_COMM_SET_NAME is provided to help debug code, it is sensible to give the same name to a communicator in all of the MPI processes where it exists, to avoid confusion. ( End of advice to users.)
The length of the name that can be stored is limited to the value of MPI_MAX_OBJECT_NAME in Fortran and MPI_MAX_OBJECT_NAME-1 in C to allow for the null terminator. Attempts to put names longer than this will result in truncation of the name. MPI_MAX_OBJECT_NAME must have a value of at least 64.

Advice to users.

Under circumstances of store exhaustion an attempt to put a name of any length could fail, therefore the value of MPI_MAX_OBJECT_NAME should be viewed only as a strict upper bound on the name length, not a guarantee that setting names of less than this length will always succeed. ( End of advice to users.)

Advice to implementors.

Implementations that pre-allocate a fixed size space for a name should use the length of that allocation as the value of MPI_MAX_OBJECT_NAME. Implementations that allocate space for the name from the heap should still define MPI_MAX_OBJECT_NAME to be a relatively small value, since the user has to allocate space for a string of up to this size when calling MPI_COMM_GET_NAME. ( End of advice to implementors.)

MPI_COMM_GET_NAME returns the last name that has previously been associated with the given communicator. The name may be set and retrieved from any language. The same name will be returned independent of the language used. comm_name should be allocated so that it can hold a resulting string of length MPI_MAX_OBJECT_NAME characters. MPI_COMM_GET_NAME returns a copy of the set name in comm_name.

In C, a null character is additionally stored at comm_name[resultlen]. The value of resultlen cannot be larger than MPI_MAX_OBJECT_NAME-1. In Fortran, comm_name is padded on the right with blank characters. The value of resultlen cannot be larger than MPI_MAX_OBJECT_NAME.

If the user has not associated a name with a communicator, or an error occurs, MPI_COMM_GET_NAME will return an empty string (all spaces in Fortran, "" in C). The three predefined communicators will have predefined names associated with them. Thus, the names of MPI_COMM_WORLD, MPI_COMM_SELF, and the communicator returned by MPI_COMM_GET_PARENT (if not MPI_COMM_NULL) will have the default of "MPI_COMM_WORLD", "MPI_COMM_SELF", and "MPI_COMM_PARENT". Passing MPI_COMM_NULL as comm will return the string "MPI_COMM_NULL". The fact that the system may have chosen to give a default name to a communicator does not prevent the user from setting a name on the same communicator; doing this removes the old name and assigns the new one.

Rationale.

We provide separate functions for setting and getting the name of a communicator, rather than simply providing a predefined attribute key for the following reasons:

• It is not, in general, possible to store a string as an attribute from Fortran.

• It is not easy to set up the delete function for a string attribute unless it is known to have been allocated from the heap.

• To make the attribute key useful additional code to call strdup is necessary. If this is not standardized then users have to write it. This is extra unneeded work that we can easily eliminate.

• The Fortran binding is not trivial to write (it will depend on details of the Fortran compilation system), and will not be portable. Therefore it should be in the library rather than in user code.

The above definition means that it is safe simply to print the string returned by MPI_COMM_GET_NAME, as it is always a valid string even if there was no name.

Note that associating a name with a communicator has no effect on the semantics of an MPI program, and will (necessarily) increase the store requirement of the program, since the names must be saved. Therefore there is no requirement that users use these functions to associate names with communicators. However debugging and profiling MPI applications may be made easier if names are associated with communicators, since the debugger or profiler should then be able to present information in a less cryptic manner. ( End of advice to users.)
The following functions are used for setting and getting names of datatypes. The constant MPI_MAX_OBJECT_NAME also applies to these names.

Named predefined datatypes have the default names of the datatype name. For example, MPI_WCHAR has the default name of "MPI_WCHAR". Passing MPI_DATATYPE_NULL as datatype will return the string "MPI_DATATYPE_NULL".

The following functions are used for setting and getting names of windows. The constant MPI_MAX_OBJECT_NAME also applies to these names.

Passing MPI_WIN_NULL as win will return the string "MPI_WIN_NULL".