Analytical Performance Models of Parallel Programs in Clusters

This paper presents a framework based on an user driven methodology to obtain analytical models on parallel systems and, in particular, clusters. This framework consists of two interconnected stages. In the first one, the analyst instruments the source code and some performance parameters are monitored. In the second one, the monitored data are used to obtain an analytical model using statistical processes. The main functionalities added to the analysis stage include an automatic fit process that provides accurate performance models and the automatic data collection from monitoring. Some examples are used to show the automatic lit process. The accuracy of the models is compared with a complexity study of the selected examples.