ABSTRACT

Modern Grids provide massive processing power and storage capacity, and demand for these resources continues to climb as researchers realize and leverage the benefits of computational science. Matching the supply of computing power and storage to the demand of users in real-time is difficult even in tightly controlled environments, but Grids with their distributed ownership and cross-domain organization add a whole new layer of complexity. As a direct result of distributed ownership, resource owners make management decision regarding local resources, often based on local information and local requirements. While this "site autonomy" is an important and even necessary property to support for Grids, these independent and myopic decisions can quickly limit the overall usefulness of the Grid - from a Grid user's perspective, resource availability can appear to fluctuate randomly, with too few resources available at times and at other times resources seemingly unused but not available for the class of user to which the user belongs. Local decisions often have unforeseen global impact. To efficiently manage the resources available in Grid systems to meet the needs of an ever-changing and diverse user community, an automated, agile, and adaptive dynamic control system with grid-wide perspective is needed. We aim to realize the many benefits of automated management in such an environment by introducing a two-tier policy architecture at the local site level and a hierarchical architecture for automated grid management that respects site autonomy. Key to our approach is the definition and enforcement of site usage policy, which defines the boundaries within which the automated system is allowed to work. Our architecture dynamically reconfigures the grid in response to user and environmental input and increases job throughput by intelligently meeting demand with the appropriate supply of computational resources. We evaluate our management architecture and local site usage policy system in several use cases that cover the classes of management actions which would be most often used in a real Grid environment. Other Recent and Upcoming Colloquia