User-Transparent Scheduling of Structured Parallel Applications in Grid Environments

The problem of scheduling parallel applications on Grids is notoriously difficult: schedulers must consider the heterogeneity of involved resources and management systems, and they often require the users to provide information about the expected application behavior. We suggest that increasingly popular structured programming approaches, using components with well-defined semantics, facilitate a more efficient and usertransparent scheduling. We illustrate our approach to scheduling using HOCs (Higher-Order Components) that capture typical patterns of parallelism (Farm-HOC, Pipeline-HOC, etc.) and come as Web Services with pre-packaged implementation and middleware setup. We take a particular Grid scheduler, KOALA, and enhance it for Grid applications built of HOCs: we propose and implement a novel scheduling algorithm that optimizes communications in such structured applications. We report experimental results of running our scheduler on the DAS-2 Grid testbed combining over 200 nodes at five sites in the Netherlands.