Thermal-Aware GreenHDFS

Data center energy costs are increasing rapidly and cooling energy costs form a substantial part of the overall energy costs. Existing research on reducing cooling costs focuses on thermal-aware task placement and migration. However, data-intensive compute clusters such as Hadoop clusters, EC2 clusters at Amazon or Google clusters, pose significant challenges to the task placement and migration algorithms. In such compute clusters, computations have strong data-locality considerations and are co-located with the data. Hence, computations cannot be arbitrarily placed on or migrated to servers in the clusters. Instead of thermal-aware task placement and migration, we take an alternate approach and focus on thermal-aware server selection for energy-aware data placement and migration. Cooling-inefficient hot spots are a ubiquitous problem in data centers. We show that by using thermally-inefficient servers as targets for energy-aware data placement, we can significantly even out the inlet thermal distribution, lower maximum exhaust temperature in the datacenter, and significantly reduce the cooling energy costs. Our technique works very well in a data-intensive compute cluster and data center simulation results with floVENT, a computational dynamics simulator, suggest a savings of 28%-40% in the overall cooling costs.NSF grant CNS 05-51665unpublishednot peer reviewedLimitedAuthor needs to update pape