Reinhard Schwarz

• In order to reduce the elapsed time of a computation, a pop-ular approach is to decompose the program into a collection of largelyindependent subtasks which are executed in parallel. Unfortunately, it isoften observed that tightly-coupled parallel programs run considerablyslower than initially expected. In this paper, a framework for the anal-ysis of parallel programs and their potential speedup is presented. Twoparameters which strongly affect the scalability of parallelism are iden-tified, namely the grain of synchronization, and the degree to which thetarget hardware is available. It is shown that for certain classes of appli-cations speedup is inherently poor, even if the program runs under theidealized conditions of perfect load balance, unbounded communicationbandwidth and negligible communication and parallelization overhead.Upper bounds are derived for the speedup that can be obtained in threedifferent types of computations. An example illustrates the main find-ings.