Tiresias: Low-Overhead Sample Based Scheduling with Task Hopping
Chunliang Hao, Jie Shen, Heng Zhang, Yanjun Wu, Mingshu Li · 2016
Sample based distributed scheduling methods have been shown to be promising lower overhead alternatives to their centralized counterparts. These methods can make fast decisions based on information gathered from just a small number of worker nodes instead of the whole cluster. Most recent works in the field tend to adopt a combination of probe actions and worker-end queues in their design. However, as individual worker nodes are becoming increasingly powerful thanks to the rapid hardware evolution, we argue that one-node sampling is now a viable choice. Specifically, we show that it is now possible to achieve even lower scheduling latency by latency by abolishing probes and worker-end queues altogether. With this insight, we introduce Tiresias, a low overhead distributed scheduler based on one-node sampling and a novel task hopping mechanism. Comparing to Sparrow's approach, experiment on Google trace shows Tiresias could reduce 20% and 60% of Sparrow's 50th percentile and 90th percentile job runtime, respectively. In addition, our experiment also shows Tiresias is especially effective in reducing the delay of small jobs in non-highly loaded clusters.