Choosing Optimal Phase Boundaries

When determining where to place the phase boundaries, the COP compiler must weigh a variety of factors. With scheduled routing, dedicating a single phase to an operator achieves the following goals:

• For a high-traffic operator, dedicating a single phase to it may be critical to achieve the necessary performance.
• Some operator implementations generate complete schedules for all the nodes (as is discussed more in Section 4.1), and therefore require a separate phase.

However, sharing a phase among a number of operators has several benefits

• It allows the operators in the phase to be used unpredictably by the application (as discussed in Section 2.4.1).
• It eliminates the overhead involved in changing phases between the operators.
• It reduces the footprint of a given set of operators in the router, thus allowing for more operators to be held in the router at once.

As a simple example, let us suppose it takes 20 cycles to switch from one phase to another for a given application and mesh. If there are two operators to schedule, the compiler can predict how long the communications for each phase will take when the operators are scheduled together or in separate phases. Let us suppose that, if scheduled in separate phases, each operator will get 0.667 bandwidth, whereas scheduled into a single phase together they will only get 0.500 bandwidth each. If each operator is used to send one message 50 words long, it will take 100 cycles per operator (200 total) if scheduled in the same phase, and 75 cycles (170 total, including 20 for phase change time) if scheduled in separate phases. Thus, the compiler will choose separate phases. If the messages were shorter (less than 40 words), a single phase holding both operators would be a better choice.