ABSTRACT
While the performance of a scalar algorithm is usually measured in terms of execution time and resource requirements, such as memory and disk usage, additional performance metrics are needed for parallel algorithms. Several factors, including communication overhead, degree of parallelism, and load imbalance, must be incorporated into parallel performance measures to be able to characterize the performance of a parallel algorithm over a wide range of parallel computers and process∗ counts. The development of a performance model capable of making realistic predictions of the parallel performance and exposing potential shortcomings of the algorithms should be an integral part of parallel program development.
