Einar Råberg Rosenvinge, Anne C. Elster, Cyril Banino
Norwegian University
of Science and Technology (NTNU)
Dept. of Computer and Info. Science (IDI)
Trondheim, Norway
email: elster@idi.ntnu.no
In this paper, results from a series of experiments with processing data-parallel MPI applications on clusters, are presented. We implement and compare several well-known scheduling strategies, and analyse various characteristics of both our experimental platforms and the applications in question, with respect to these scheduling strategies.
More precisely, we take into account cluster bandwidth and latency, the (possibly different) processor speeds, the cluster's I/O capabilities, as well as the application's communication to computation ratio and algorithmic variance.
Based on these characteristics, we aim at determining how to pick the best scheduling strategy, i.e. the one which minimizes the total execution time, and verify our results through implementation.