Taming Completion Times in a Cloud Cluster
Michael Wang Bharath Balasubramanian Yu Xiang Tian Lan and Mung Chiang
We present a framework to leverage job heterogeneity with respect to the client’s tolerance for service latency. Specifically, we model job utilities as a function of job completion times, and formulate a problem to maximize the overall utility achieved by all jobs in a cluster. To solve this problem, we present the TACT algorithm that provides a solution with a bounded optimality gap. Finally, we validate our method on the Hadoop framework, and compare our results favorably with common policies for resource allocation.