Predicting the speedup of parallel Ada programs

Abstract

A method for predicting the speedup of parallel Ada programs has been developed. For this purpose the states Active and Blocked are used to characterize a task during program execution. Of the active tasks, some may be waiting for a processor to be available. Transitions between the two states may be caused only by certain tasking constructs that can be statically identified in the source code. The execution of a task forms a list of Active and Blocked time segments. Segments in different tasks may depend on each other through task synchronizations, thus forming a dependency graph.

Using this graph and certain assumptions about the way tasks are scheduled, one can determine how the number of active tasks varies during the execution. Disregarding hardware and system overheads, speedup is limited either by the number of processors or by the number of active tasks. That is, dependency graphs make it possible to compare the speedup of different programs solving the same problem. This method can also be used for selecting a multiprocessor system with a suitable number of processors for a certain program.

By inserting probes at certain tasking constructs and executing the program on a single-processor, we are able to record the dependency graph. This method has been used for predicting the speedup of a parallel Ada program containing 80 tasks.