On the integration of model-based performance optimization and program implementation

The author describes an approach to integrate system design and implementation of embedded parallel systems into a programming environment. The computation of the design decisions is based on the branch-and-bound-with-underestimate technique, previously applied to compute the optimal schedule for precedence constraint task graphs in the presence of communication cost. The authors have extended this method to compute the optimal system configuration (i.e., the configuration which yields optimal performance) consisting of the network topology, the schedule, and the placement of buffers. In addition, a quasi-static scheduling strategy to handle input-dependent run-time behavior is described. The ideas presented here have been implemented in the Data Stream Processing Language (DSPL) programming environment, which consists of a programming language and an integrated set of tools supporting automated system design as well as implementation