Speculation-aware Resource Allocation for Cluster Schedulers

Resource allocation and straggler mitigation (via “speculative” copies) are two key building blocks for analytics frameworks. Today, the two solutions are largely decoupled from each other, losing the opportunities of joint optimization. Resource allocation across jobs assumes that each job runs a fixed set of tasks, ignoring their need to dynamically run speculative copies for stragglers. Consequently, straggler mitigation solutions are unsure of how to provide for speculative tasks. In this thesis, we propose Hopper, a new speculationaware job scheduler. Hopper dynamically allocates resources for effective speculation based on theoretically derived guidelines that provably minimize average job completion time. In addition, Hopper also provides a simple knob for trading off fairness and performance. Evaluations on a 200-node cluster with Hadoop and Spark prototypes, on production workloads from Facebook and Microsoft Bing, show job speedups of 50% − 70% compared to SRPT schedulers, fairness schedulers, and straggler mitigation algorithms.1 The paper is undersubmission to 12th USENIX Symposium on Networked Systems Design and Implementation (NSDI ’15).