Network scheduling for computational grid environments

The problem of data movement is entral to distributed omputing paradigms like the Grid. While often overlooked, the time to stage data and binaries an be a signi ant ontributor to the wall- lo k program exe ution time in urrent Grid environments. This paper des ribes a simple s heduler for network data movement in Grid systems that an adaptively determine data distribution s hedules at runtime on the basis of Network Weather Servi e (NWS) performan e predi tions. These s hedules take the form of spanning trees. The distribution me hanism is an enhan ement to the Logisti al Session Layer (LSL), a system for optimizing data transfers using logisti s. Key words. Grid omputing, data logisti s, data staging 1. Introdu tion. As Computational Grid environments proliferate, the ommunity must onstantly evolve the way in whi h omputing systems are used. Distributed omputing on the Grid has enabled new ways of harnessing omputing resour es and yet, has exposed its own set of hallenges. One su h problem is that of data movement. Appli ations that are drawn to the Grid be ause of large resour e requirements frequently onsume or generate large amounts of data. The problems of data lo ality and data movement are be oming more prominent and riti al to the performan e and deployability of Grid systems. Further, due to the dynamism inherent in Grid environments, it is lear that me hanisms for data staging must be adaptive like the omputations themselves