Generalised Rate Control Approach For Stream Processing Applications

TL;DR

This paper introduces a graph neural network-based deep reinforcement learning method for rate control in distributed stream processing, improving system stability and performance by proactively managing data emission rates.

Contribution

It presents a novel GNN-based reinforcement learning approach for rate control that adapts to multiple scenarios and outperforms traditional methods in stream processing systems.

Findings

Up to 13.5% increase in throughput

Up to 30% reduction in latency

Effective adaptation across multiple applications

Abstract

Distributed stream processing systems are widely deployed to process real-time data generated by various devices, such as sensors and software systems. A key challenge in the system is overloading, which leads to an unstable system status and consumes additional system resources. In this paper, we use a graph neural network-based deep reinforcement learning to collaboratively control the data emission rate at which the data is generated in the stream source to proactively avoid overloading scenarios. Instead of using a traditional multi-layer perceptron-styled network to control the rate, the graph neural network is used to process system metrics collected from the stream processing engine. Consequently, the learning agent (i) avoids storing past states where previous actions may affect the current state, (ii) is without waiting a long interval until the current action has been fully effective and reflected in the system's specific metrics, and more importantly, (iii) is able to adapt multiple stream applications in multiple scenarios. We deploy the rate control approach on three applications, and the experimental results demonstrate that the throughput and end-to-end latency are improved by up to 13.5% and 30%, respectively.