Reparallelization and migration of OpenMP applications in grid environments

Grid computing opens up a new computing infrastructure and users gain access to a vast landscape of computing resources that are spread over the world and that are instantly accessible through the Internet. In the future, computational Grids are expected to deliver high-end computing power at the user’s finger tips. Today, typical users of a computational Grid only exploit a single cluster and are faced with the job schedulers that assign computing resources to applications. Job schedulers expect a resource estimation to exclusively allocate parts of the computing system for the application. Users typically over-estimate the resource limits of their applications to avoid premature termination of the application by the job scheduler. Over-estimating resource limits has negative effects on both the clusters’ schedules as well as on the waiting time until an application eventually executes. In this work, we present a solution that alleviates the need of estimating the resource limits for OpenMP applications. A reparallelization of OpenMP applications is automatically computed, when new computing resources become available or are withdrawn. The application can be migrated between clusters of the Grid if an allocated resource is about to be exceeded. Programmers do not need to change the application to enable reparallelization and migration. Instead, our compiler transparently prepares the applications for reparallelization and migration. It adds code to the application to enable reparallelization and augments the application with a platformindependent, coordinated checkpointing algorithm for migration. A prototype implementation of a migration framework automatically discovers free resources and migrates the application to these resources. Measurements show that the overhead of reparallelization and migration is roughly 4%, which we consider a negligible cost compared to the gain of flexibility.