Controlled eager evaluation in a Dynamic-Arc Tagged-Token Dataflow Model

The data-driven execution scheme achieves parallelism by simultaneously activating all operations whose data is available. However, in a dataflow machine with limited resources, the data- driven scheme can lose efficiency when a significant amount of time is spent on the execution of useless computations. In this paper, we introduce a general dataflow model, the Dynamic-Arc Tagged-Token Dataflow Model, which extends traditional dataflow models to include dynamic arcs as a natural mechanism to implement function and procedure calls. Within this model, we develop the Efficient Data-Driven Evaluation Scheme which improves the performance of conventional data-driven execution by terminating useless computations at run-time. This is facilitated by the construction of an extended dataflow graph at compile-time in which useless computations can be identified dynamically. When useless computations are terminated, program execution times improve since more resources are available for the processing of computations which contribute to the final results. 12 refs., 9 figs.