Agents, clusters and components: A synergistic approach to the GSP

Grids provide access to a vast amount of computational resources for the execution of demanding computations. These resources are geographically distributed, owned by different organizations and are vastly heterogeneous. The aforementioned factors introduce uncertainty in all phases of a Grid Scheduling Process (GSP). This work describes a synergistic multidisciplinary approach which aims at addressing this uncertainty. It proposes a network of resource representatives (RRs), which maintain the more or less static characteristics of available workers they represent. Clustering techniques are used for the efficient searching in the network of RRs by client agents. After the discovery of possibly suitable resources, client agents and resource agents negotiate directly for the selection of the best available resource set. Finally, according to the characteristics of the selected resource set and its current state, we propose a component-based application configuration approach based on component variants, that adjusts the application for the forthcoming execution phase in the selected resource set. We evaluate our approach using simulation and we show that it outperforms centralized index approaches for large computational grids.

► Resource representatives (RRs) maintain the static properties of grid workers.
► Clustering techniques support searching of the network of RRs by client agents.
► Client agents optimize selection by pre-execution negotiation with resource agents.
► Component-based configuration, adjusts applications to the selected resource set.
► Evaluation of the approach shows that it outperforms centralized index approaches.