50 likes | 206 Views
Three Pieces of the MapReduce Workload Management Puzzle. Abhishek Verma 1,2 , Lucy Cherkasova 2 , Roy H. Campbell 1 1 University of Illinois at Urbana-Champaign, 2 HP Labs. Motivation. Often MapReduce applications part of critical business pipelines
E N D
Three Pieces of the MapReduce Workload Management Puzzle Abhishek Verma1,2, Lucy Cherkasova2, Roy H. Campbell1 1University of Illinois at Urbana-Champaign, 2HP Labs SOSP 2011
Motivation • Often MapReduce applications part of critical business pipelines • Require job completion time guarantees (SLOs) • Existing MapReduce schedulers do not support Service Level Objectives • Goal: Design a workload management framework for MapReduce jobs with completion time goals in shared environments
Three Pieces of the Puzzle Job Ordering How much resources? Allocating spare resources
Job Scheduling with Different Mechanisms • Earliest Deadline First • Does not require any job information • Min-EDF • Automatically extract job profiles from past executions • Compute and allocate minimum resources • Min-EDF-WC • Allocate any spare resources among running jobs • When new job arrives, compute if enough slots will be released in the future to satisfy current job • If not, cancel spare tasks of currently running jobs
Evaluation Min-EDF-WC leads to smaller job completion times than Min-EDF Min-EDF-WC misses 2 times lesser job deadlines than Min-EDF