logical messages into a single physical message, and to incur the startup cost only once. MPI's derived data types can be helpful in such packing and unpacking tasks.
8.4.4 Intercommunicators
Normal communicators (see Section 8.2.1) are constructed by elimination. That is, one begins with MPI_COMM_WORLD, the group of all active processes, and communicators are derived by selecting subsets from this communicator or its descendents. Objects constructed this way can only become smaller, never larger. Occasionally one finds that this logical structure is insufficient, and it is desirable to construct a new communicator from the union of two existing communicators. These are called inter-communicators, and are used primarily in applications that display functional parallelism.
8.4.5 MPI-2
The MPI-2 standard was released in April 1997, and defines support for several new features. The most important of these are:
Process creation and management MPI-1 programs consist of processes that are statically allocated at runtime and that exist through the life of the program. MPI-2 also allows processes to be dynamically created and destroyed. Much of this functionality is derived from the extensive experience with the PVM research effort.
One-sided communications In some dynamic applications it can be useful for either the sender or receiver of a communication to specify all the parameters of the transfer, i.e., both the source and destination. This relieves the other party of needing to know the identity or timing of requests or replies. There is also the possibility of exploiting very fast communication mechanisms such as the direct memory access capabilities that exist on shared memory processors.
Extended Collective Operations MPI-1 does not allow for collective operations involving intercommunicators. This restriction is relaxed in MPI-2, along with the introduction of new procedures for the creation of intercommunicators, and two new collective operations.
External interfaces Features are provided so that users and library designers can access the internals of MPI's opaque objects. This allows for new functionality to be layered on top of MPI's internal data structures. The relationship between MPI and generic thread libraries is also defined.
