As clusters and grids become increasingly popular alternatives to custom-built parallel computers, they expose a growing heterogeneity in processing and communication capabilities. Performing an effective load balancing on such environments can be best achieved when the heterogeneity factor is quantified and appropriately fed into the load balancing routines. In this work, we discuss an approach based on constructing a tree model that encapsulates the topology and the capabilities of the cluster. The different components of the execution environment are dynamically monitored and their processing and communication capabilities are collected and aggregated in a simple form easily usable when load balancing is invoked. We used the model, called DRUM, to guide load balancing in the adaptive solution of three numerical problems on various cluster configurations. The results showed a clear benefit from using DRUM, in clusters with computational or network heterogeneity as well as in non-dedicated clusters. The results also showed that DRUM generates a very low overhead.
|Citation (BIBTEX)||Full Thesis (PDF; 1.08 MB)|