A scalable and low latency probe-based scheduler for data analytics frameworks

摘要

Today's data analytics frameworks divide jobs into many parallel tasks such that each task operates on a small partition of data in order to execute jobs with low latency. Such frameworks often rely on probe based distributed schedulers to tackle the challenge of reducing the associated overhead. Unfortunately, the existing solutions do not perform efficiently under workload fluctuations and heterogeneous job durations. This is due to a problem called Head-of-Line blocking, i.e., short tasks are enqueued at workers behind longer tasks. To overcome this problem, we propose Peacock (Khelghatdoust and Gramoli, 0000) [25] a new fully distributed probe-based scheduling method. Unlike the existing methods, Peacock introduces a novel probe rotation technique. Workers form a ring overlay network and rotate probes using elastic queues of workers. It is augmented by a novel starvation-free probe reordering algorithm executed by workers. We evaluate Peacock against two existing state-of-the-art probe based solutions through a trace driven simulation of up to 20,000 workers and a distributed experiment of 100 workers in Apache Spark under Google, Cloudera, and Yahoo! traces. The performance results indicate that Peacock outperforms the state-of-the-art in all cluster sizes and loads. Our distributed experiments confirm our simulation results.