MetaLoRaS: A Re-scheduling and Prediction MetaScheduler for Non-dedicated Multiclusters
García, J. R.
Hernández Budé, Porfidio
MetadataShow full item record
Recently, Multicluster environments have become more important in the high-performance computing world. However, less attention has been paid to non-dedicated Multiclusters. We are developing MetaLoRaS, an efficient twolevel MetaScheduler for non-dedicated environments, which assigns PVM and MPI applications according to an estimation of the turnaround time in each particular cluster. The main MetaScheduler goal is to minimize the average job turnaround time in a non-dedicated environment. The efficiency of MetaLoRaS depends on the prediction accuracy of the system and its ability to take decisions according to changes in local workload. In this paper we present different Metascheduling techniques that take the dynamics of the local workload into account and compare their effects on system performance. We evaluate the prediction accuracy in relation to the low-level queues sizes. Finally, we analyze the relationship between prediction accuracy and system performance.