A parallel task allocation methodology for non-buffered message-passing environments

The performance of large-scale real-world applications may be enhanced by more efficient utilization of today´s powerful parallel computing hardware. The authors describe a novel MIMD (multiple instruction, multiple data) static task allocation methodology created to achieve this goal for a wide range of real-time and deterministic systems. The method performs automatic assignment, mapping, and scheduling of the executable tasks to the available set of processors, assuming synchronous nonbuffered (lock-step) communication between the sending and receiving processors and an arbitrary static message-passing topology. The algorithm incorporates a set of list-based heuristics and graph-theoretical procedures designed to balance computational load with communication requirements. The effectiveness of the method has been verified by applying it to a large number of randomly generated task systems that span a wide range of inherent concurrency as well as to a task system of a real-world simulation of the Space Shuttle main rocket engine