Mary, Hugo, and Hugo*: Learning to schedule distributed data‐parallel processing jobs on shared clusters

Abstract

Summary Distributed data‐parallel processing systems like MapReduce, Spark, and Flink are popular for analyzing large datasets using cluster resources. Resource management systems like YARN or Mesos in turn allow multiple data‐parallel processing jobs to share cluster resources in temporary containers. Often, the containers do not isolate resource usage to achieve high degrees of overall resource utilization despite overprovisioning and the often fluctuating utilization of specific jobs. However, some combinations of jobs utilize resources better and interfere less with each other when running on the same shared nodes than others. This article presents an approach for improving the resource utilization and job throughput when scheduling recurring distributed data‐parallel processing jobs in shared clusters. The approach is based on reinforcement learning and a measure of co‐location goodness to have cluster schedulers learn over time which jobs are best executed together on shared resources. We evaluated this approach over the last years with three prototype schedulers that build on each other: Mary, Hugo, and Hugo*. For the evaluation we used exemplary Flink and Spark jobs from different application domains and clusters of commodity nodes managed by YARN. The results of these experiments show that our approach can increase resource utilization and job throughput significantly.

Collaborative Cluster Configuration for Distributed Data-Parallel Processing: A Research Overview

“…Additional challenges are introduced when jobs do not use resources in isolation, but share access and potentially interfere with each other, impeding individual job performance often significantly [29,27].…”

Tarema: Adaptive Resource Allocation for Scalable Scientific Workflows in Heterogeneous Clusters

“…Since SJFN assigns all tasks to the most powerful machines, many tasks have to share the resources. This sharing can lead to interferences, which can get higher with a higher number of competing tasks [41]- [43].…”

On the Future of Cloud Engineering

“…Many other works apply reinforcement learning to integrate the exploration of potential solution spaces directly with an optimization towards given objectives such as high resource utilization, low interference, and cluster throughput. In this way, several novel cluster schedulers use either classical or deep reinforcement learning methods to schedule various types of cluster jobs in large data center infrastructures [127]- [129]. Other systems use reinforcement learning, for example, to re-provision and scale microservices towards given service-level objectives [130].…”