| The grid is emerging as the platform for resource intensive applications. A grid middleware service, Application Scheduling, maps applications to resources that best match them. Resource factors (availability, distribution, load, access policies) and application factors (distribution, resource requirements) influence scheduling decisions. Due to the large scale of the grid; tracking all resource state changes is unfeasible. Thus, distributed concurrent schedulers have to rely on dynamic and potentially inconsistent resource information. In this thesis, we propose using probabilistic forwarding protocols--- Non-uniform protocols---to pull resource information in real time. We compare this reactive approach with the proactive non-uniform resource information push approach considering their impact on scheduling quality and success. We also propose selection criteria to choose the best matched resources. We focus on three selection criteria: resource distance from requester, past behavior; and resource information freshness. (Abstract shortened by UMI.)... |
