PCSsampler: Sample-based, Private-state Cluster Scheduling

Chunliang Hao, Jie Shen, Celia Chen, Heng Zhang, Yanjun Wu, Mingshu Li

Abstract: As a promising alternative to centralized scheduling, sample-based scheduling is especially suitable for high fan-out workloads that contain a large number of interactive jobs. Compared to centralized schedulers, existing sample-based schedulers do not hold a global view of the cluster's resource status. Instead, the scheduling decisions are made solely based on the status of a small set of randomly sampled workers. Although this simple approach is highly efficient in large clusters, the lack of global knowledge of the cluster can lead to sub-optimal task placement decisions and difficulties in enforcing global scheduling policies. In this paper, we address these challenges in existing sample-based scheduling approaches by allowing the scheduler to maintain an approximate version of the global resource status through caching the worker node's status extracted from reply messages. More specifically, we introduce the private cluster-state technique (PCS) for the scheduler to obtain such global information. We show that the scheduler can make better scheduling decisions by utilizing PCS and the scheduler can become more capable in enforcing global scheduling policies. The use of PCS is of low cost since it does not initiate new communication in sample-based scheduling. Our approach is implemented in PSCSampler, a full distribute sample-based scheduler, which gains global knowledge from PCS. Experiment results from both simulation runs and Amazon cluster runs show that compared to Sparrow, PCSsampler can significantly reduce both 50 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> percentile and 90 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> percentile runtime. The firsttime success rate of PCSsampler in gang scheduling is closer to an omniscient centralized scheduler than baseline sample based scheduler.