Distributed Decision Making for Metaschedulers

We propose a distributed scheduling algorithm for HPC grids and clouds that allows users, resource providers, and grid community to participate in the scheduling decision. Its hierarchical representation of dynamically configurable criteria integrates arbitrary aspects into the process in a well-defined manner; thus, the algorithm can accommodate different views and policies of geographically distributed stakeholders. The heuristic permits resource providers to exactly define the share of participation allocated to each party and to retain control over their resources. It is based on the Analytic Hierarchy Process but has been extended to support the dynamic evaluation of utility values. Evaluation respects the different perception of utility values by human decision makers with regard to their dissemination on the measure scale. The algorithm is incorporated into a distributed metascheduling architecture that is modeled around cooperating peers. Its design is resilient to network and site failures and preserves the autonomy of independent providers.