A Deep Dive into the Google Cluster Workload Traces: Analyzing the Application Failure Characteristics and User Behaviors

TL;DR

This paper analyzes Google's large-scale cloud workload traces to understand failure patterns and user behaviors, aiming to improve failure prediction and resource utilization in data centers.

Contribution

It provides a comprehensive analysis of failure characteristics and user heterogeneity in Google Cluster traces, proposing insights for early failure prediction systems.

Findings

Failed jobs have identifiable resource and scheduling patterns.

Certain users dominate job submission and failure events.

Insights can inform dynamic rescheduling to reduce failures.

Abstract

Large-scale cloud data centers have gained popularity due to their high availability, rapid elasticity, scalability, and low cost. However, current data centers continue to have high failure rates due to the lack of proper resource utilization and early failure detection. To maximize resource efficiency and reduce failure rates in large-scale cloud data centers, it is crucial to understand the workload and failure characteristics. In this paper, we perform a deep analysis of the 2019 Google Cluster Trace Dataset, which contains 2.4TiB of workload traces from eight different clusters around the world. We explore the characteristics of failed and killed jobs in Google's production cloud and attempt to correlate them with key attributes such as resource usage, job priority, scheduling class, job duration, and the number of task resubmissions. Our analysis reveals several important characteristics of failed jobs that contribute to job failure and hence, could be used for developing an early failure prediction system. Also, we present a novel usage analysis to identify heterogeneity in jobs and tasks submitted by users. We are able to identify specific users who control more than half of all collection events on a single cluster. We contend that these characteristics could be useful in developing an early job failure prediction system that could be utilized for dynamic rescheduling of the job scheduler and thus improving resource utilization in large-scale cloud data centers while reducing failure rates.