Neural-based Modeling for Performance Tuning of Spark Data Analytics

TL;DR

This paper introduces a deep learning-based approach for performance modeling of Spark data analytics workloads, enabling more accurate predictions for system tuning in cloud environments.

Contribution

It proposes workload embedding techniques and evaluates various deep learning models, demonstrating superior performance over existing tools for cloud analytics.

Findings

Deep learning models outperform traditional performance models.

Workload embeddings effectively capture computational characteristics.

The best model shows significant accuracy improvements.

Abstract

Cloud data analytics has become an integral part of enterprise business operations for data-driven insight discovery. Performance modeling of cloud data analytics is crucial for performance tuning and other critical operations in the cloud. Traditional modeling techniques fail to adapt to the high degree of diversity in workloads and system behaviors in this domain. In this paper, we bring recent Deep Learning techniques to bear on the process of automated performance modeling of cloud data analytics, with a focus on Spark data analytics as representative workloads. At the core of our work is the notion of learning workload embeddings (with a set of desired properties) to represent fundamental computational characteristics of different jobs, which enable performance prediction when used together with job configurations that control resource allocation and other system knobs. Our work provides an in-depth study of different modeling choices that suit our requirements. Results of extensive experiments reveal the strengths and limitations of different modeling methods, as well as superior performance of our best performing method over a state-of-the-art modeling tool for cloud analytics.