Smart Scheduling of Continuous Data-Intensive Workflows with Machine Learning Triggered Execution

TL;DR

This paper presents SmartFlux, a machine learning-based middleware that optimizes continuous data workflows by adaptively triggering computations, significantly reducing resource usage while maintaining output accuracy within acceptable error bounds.

Contribution

It introduces a novel workflow model that relaxes trigger conditions based on input-output impact, leveraging machine learning to improve efficiency in data-intensive processing.

Findings

Resource savings of up to 40% without significant output deviation

High confidence in maintaining output accuracy within error bounds

Seamless integration with existing workflow managers

Abstract

To extract value from evergrowing volumes of data, coming from a number of different sources, and to drive decision making, organizations frequently resort to the composition of data processing workflows, since they are expressive, flexible, and scalable. The typical workflow model enforces strict temporal synchronization across processing steps without accounting the actual effect of intermediate computations on the final workflow output. However, this is not the most desirable behavior in a multitude of scenarios. We identify a class of applications for continuous data processing where workflow output changes slowly and without great significance in a short-to-medium time window, thus wasting compute resources and energy with current approaches. To overcome such inefficiency, we introduce a novel workflow model, for continuous and data-intensive processing, capable of relaxing triggering semantics according to the impact input data is assessed to have on changing the workflow output. To assess this impact, learn the correlation between input and output variation, and guarantee correctness within a given tolerated error constant, we rely on Machine Learning. The functionality of this workflow model is implemented in SmartFlux, a middleware framework which can be effortlessly integrated with existing workflow managers. Experimental results indicate we are able to save a significant amount of resources while not deviating the workflow output beyond a small error constant with high confidence level.