Improved scheduling of signal flow graphs onto multiprocessor systems through an accurate network modelling technique

This paper presents a new integrated technique for accurately scheduling acyclic precedence expansion graphs (APEGs) onto multiprocessor networks with various topologies. APEGs specify the nature, connectivity, and precedence relationships of all tasks, as well as data amounts that pass between tasks. We present a scheduling algorithm that uses the Branch and Bound technique when the number of tasks in the graph is small, and the Earliest Task First heuristic, otherwise. Graph partitioning and scheduling algorithms however, require a good estimate of interprocessor communication (IPC) costs within the target network. We thus use a technique called Successive Superposition to accurately determine IPC costs. Successive Superposition provides a methodology to decompose a complex network, containing primarily deterministic traffic, into simpler queueing models which may then be analyzed in isolation. By combining a heuristic scheduling algorithm with this exact IPC analysis technique, we avoid unpredictable behavior and incorrect mapping decisions that could result in longer graph execution times. We present an example in which an inaccurate assessment of IPC costs leads to a 22% to 30% increase in the graph execution time.