SCALING PARALLEL PROGRAMS FOR MULTIPROCESSORS - METHODOLOGY AND EXAMPLES

To design effective large-scale multiprocessors, computer architects need to understand how parallel applications will be scaled to use these machines. Previous studies of scaling have defined useful models for the constraints under which parallel applications might be scaled, but they do not show how applications might actually be scaled to meet these constraints. Typically, data set size is the only parameter scaled. The authors argue that scaling real applications is more complicated. In scientific applications that simulate physical phenomena, for example, a scientist will want to scale other application parameters in conjunction with the data set size to obtain useful results. Realistic scaling methods for these applications are usually based on the contributions that different application parameters make to the error in the simulation. This article uses the widely applicable principle of scaling parameters so that their error contributions are about equal. Through examples, the authors show that scaling in realistic ways can lead to different conclusions about the effectiveness and design of larger multiprocessors. The authors also conclude that system designers must understand the relevant characteristics of the benchmark applications they use if they are to arrive at correct design decisions for building large multiprocessor systems.