Jim Blythe ISI, USC AI-based approaches to workflow management in Grids --------------------------------------------------- Abstract Grid computing promises the ability for users to harness the power of large numbers of heterogeneous, distributed resources, including computing resources, data storage systems and instruments. The vision is to enable users and applications to seamlessly access these resources to solve complex large-scale problems. Data intensive scientific applications are no longer being developed as monolithic bodies of code. Instead, standalone application components are combined to process the data in various ways. The applications can now be viewed as complex workflows, consisting of various transformations performed on the data. These workflows often require the Grid to execute efficiently because of the large amounts of computation and data involved. In current environments, users need to discover resources manually and schedule the jobs directly onto the Grid, essentially composing detailed workflow descriptions by hand. This leaves users struggling with the complexity of the Grid and weighing which resources to use, where to run the computations, where to access the data etc. We are working to automate this workflow generation process as far as possible and to provide knowledge-based assistance to users where needed. We describe an architecture that integrates several grid services to take a high-level workflow specification, locate appropriate resources and oversee its execution on the Grid. Within this architecture we use AI planning to generate workflows from logical descriptions of the required data. The system has been used to execute gravitational physics workflows involving hundreds of individual tasks. Some of the challenges for planning in Grids include (1) how to distribute the knowledge and the reasoning processes for workflow generation across the Grid and (2) monitoring and repairing workflows as the environment changes during their execution. This is joint work with Yolanda Gil, Ewa Deelman and others at ISI. Bio: Jim Blythe is a computer scientist in the Information Sciences Institute at USC. His interests include planning, knowledge acquisition and intelligent user interfaces. He is currently working on planning for Grids and on acquiring advice for planning systems from end users. He is chairing a workshop this year on planning and scheduling for web and grid services at the international planning conference, and is a manager of the Global Grid Forum research group on Workflows. He received his PhD from Carnegie Mellon University in 1998 on planning under uncertainty.