Modes.

Design of such a library requires tradeoffs among convenience, efficiency, and scalability. We prefer to allow he user to make such decisions by providing multiple modes in which the processes created and controlled by the user interact with the X Window server. They are:

Independent

Each process connects directly to the X Window server and my issue its own individual drawing commands.

Shared-Server

Some processes connect; all may draw. The MPE Graphics Library will deliver the operations to the server internally.

Multi-Server

Some processes connect and are dedicated to drawing. This generalizes what is perhaps the most commonly used way of obtaining graphical output from parallel programs, in which one process is responsible from interacting with the X Window server.

Collective

Some processes connect, and drawing operations are collective (that is, the drawing operations have the same usage semantics as the MPI collective routines).

All drawing routines may be used in any of the four modes (with restrictions noted below). By using the same interface for all modes, the programmer can easily modify the tradeoffs for scalability and response time without changing the details of the individual drawing commands.

The independent mode is easily understood. In the Multi-Server mode, the MPE_Open_graphics call selects some nodes as servers; these make a connection to the graphics display. An MPI communicator that contains only those server nodes is created. Subsequent graphics calls from non-server processes send messages to some server process; graphics calls on a server process both make Xlib calls and check for incoming graphics requests from non-server nodes. Note that for the Multi-Server mode to work, we need the contexts of MPI, since pending server requests must be kept independent from operations such as MPI_Recv( ..., tag = MPI_ANY_TAG, source = MPI_ANY_SOURCE, ... ) .

The Shared-Server mode is like the Multi-Server mode, except that the server nodes only act as graphics servers. In Collective mode, drawing operations and input operations are synchronous and collective. As in other MPI collective operations, all processes in the associated communicator must make the call.