Conclusions
	l Clusters are sensitive to small policy changes – Clusters are built for specific workloads – Their performance is important and small changes have significant impact l No scheme is universally applicable l Future Work – Real system implementation • Real Workloads • Real node level resource management – More steady performance