Dynamic Load Balancing for Solution Adaptive FEA on Heterogenious Clustures

Sponsored by: National Science Foundation
Project period: 1996

Being able to use as much of the available compute cycles in a network as possible is an area of active research at the present time. A suitable load balancing strategy is essential if efficient use of the compute servers on a given network is to be assured. On heterogeneous networks of workstations, widely differing processing capabilities coupled with unpredictable variations in work load can pose a serious load balancing problem. In addition, adaptive finite element analyses add further complexities by generating dynamically varying amount of work load in different parts of the computational domain as the solution evolves.

In distributed finite element analysis, the load balancing issues have been addressed, typically using domain decomposition methods. However, most of these strategies are static (or pseudo-static) and rely on knowledge of system load and computational complexity of the numerical problem. In case of solution adaptive finite element analysis on a heterogeneous network, such assumptions are clearly violated leading to poor computational efficiency.

The current work addresses the above-mentioned issues by classifying the imbalances using a severity measure and deals with each class of imbalance separately. It uses concepts like buffer zones and migration to maintain the computation at a near-optimal load balanced state. The strategies are adopted dynamically and are naturally suitable for networks that may contain computers with widely varying capabilities and loads.



Questions, comments or problems regarding this service? webmaster@tx.altair.com.

Copyright 1994 Computational Mechanics Company Inc.