Multiple Threads of Control

  One important feature that the language should be able to express is multiple threads of control. In other words, the language should be able to represent a more MIMD model with different regions of the mesh executing different parts of the application. (Multiple threads on a single node are orthogonal to this issue, and their communication needs must be expressed using the multiple-operator methods of Section 2.4.1.)

Given a set of operators A0, A1, ... that are active on the left side of a mesh, and a similar set B0, B1, ... active on the right side, the concept of multiple threads of control can be represented as follows:

The result will be two threads of control, one looping through the A operators on the left, one through the B operators on the right. The application thus has multiple phases that are divided not just in time but in space as well. The spatial extent of a phase is the region of the mesh in which communication and computation are occurring for the operators in that phase. In the above example, the A operators' spatial extent is the left side of the mesh.

For dynamic routers, the spatial extent is largely irrelevant for communications, though the compiler will assume the communications are mostly non-overlapping when it attempts to find efficient implementations of the operators on either side. For scheduled routers, multiple simultaneous phases implies that, rather than computing fully global schedules, ``global'' schedules are computed for subsets of the mesh independently. The issue of extents is discussed in detail in Section 6.1.