Multiple Job Allocation in Multicluster System
Hector Blanco (Universitat de Lleida)
Damia Castella (Universitat de Lleida)
Josep Lluis Lerida (Universitat de Lleida)
Fernando Guirado (Universitat de Lleida)
In multi-cluster systems, job scheduling is delegated to a meta-scheduler that has a global vision of the resources. With this knowledge it determines the best execution cluster to allocate the parallel jobs to. In some cases, resources from more than one cluster could be needed to execute these. This situation, called co-allocation, must ensure that job execution does not affect others present in the shared clusters. Obtaining a lower execution time is a big challenge, even more so when there is an execution queue with multiple jobs in the system. In this work, the authors present a linear programming model that allows simultaneous scheduling of all the parallel jobs in the system queue, instead of allocating them individually. This model minimizes the average execution time for all of these, and is capable of taking advantage of co-allocation avoiding the saturation of the network links.
Parallel and Distributed Computing