Maximizing utilization has been the approach for shared homogeneous clusters to achieve short job completion times (hence policies such as join-shortest-queue, least-work-left, etc.) With computing co-located with data, as in Hadoop / MapReduce clusters, jobs are scheduled, instead, to maximize data locality. But is utilization, or data locality, the right metric in this case?
Yi Lu is an assistant professor in the ECE department at UIUC. She is a recipient of the Stanford Graduate Fellowship, the Sigmetrics Best paper award in 2008, the Performance Best paper award in 2011 and the NSF Career award in 2012. Her work focuses on developing scalable architectures and algorithms for large networking systems such as modern web services with dynamic content, cloud computing, and social networks. Her work spans fundamental analysis and algorithm implementation, and emphasizes design of low-complexity distributed algorithms.
Faculty Host: Mor Harchol-Balter
harchol [atsymbol] cs.cmu.edu