A Self-Adapting Task Scheduling Algorithm for Container Cloud Using Learning Automata

Lilu Zhu, Kai Huang, Yanfeng Hu, Xianqing TaiIEEE Access 2021 · DOI

Abstract

In dynamic container cloud environments, maintaining resource balance is a critical challenge for ensuring service quality. Traditional static and heuristic-based scheduling methods often struggle with the high volatility of containerized workloads. This paper presents ADATSA, a self-adapting task scheduling algorithm based on Learning Automata (LA). By employing a novel reward-penalty mechanism that considers node priority, node stability, and task migration history, ADATSA effectively navigates environmental uncertainty. Experimental results on a 53-node Kubernetes cluster demonstrate that ADATSA reduces resource imbalance by up to 23.24% and improves throughput by over 6% compared to standard Kubernetes and heuristic schedulers.

Introduction

The rise of container orchestration platforms like Kubernetes has enabled unprecedented scalability and flexibility for cloud-native applications. However, the dynamic nature of these environments—characterized by frequent scaling, updates, and failures—often leads to significant resource fragmentation and imbalance. When some nodes are over-utilized while others remain idle, the overall Quality of Service (QoS) suffers. This paper addresses the limitations of existing scheduling approaches by proposing ADATSA, an intelligent, adaptive scheduling framework.

Problem Statement

Standard schedulers often rely on a "one-and-done" placement strategy. Once a task is assigned to a node, it remains there unless a failure occurs or a manual trigger is pulled. This static approach is ill-suited for the fluctuating resource demands of modern microservices. While dynamic scheduling methods like Particle Swarm Optimization (PSO) exist, they are often computationally intensive. Furthermore, simple rule-based rebalancing can lead to "ping-pong" effects, where tasks are repeatedly migrated, causing system instability. There is a clear need for a scheduling model that can learn from the environment and make granular, stability-aware decisions.

Proposed Methodology: ADATSA and TSOLAM

The core of the proposed solution is the Task Scheduling Oriented Learning Automaton Model (TSOLAM). This model treats the scheduling process as an interaction between an automated agent and a stochastic environment.

The interaction between Learning Automata and the Stochastic Environment.

The ADATSA algorithm utilizes a specialized Reward-Penalty (RP) strategy. Unlike basic LA models that use binary feedback, ADATSA calculates a Reward Base Probability based on three key metrics:

  • Node Priority: A measure of a node's current resource availability.
  • Node Stability: A historical record of how reliably a node hosts tasks.
  • Task Stability: A metric tracking how often a specific task has been migrated, preventing the system from over-migrating "restless" containers.

The system is supported by an Environment & Load Monitor (ELM), which identifies high-load tasks and places them in a prioritized buffer queue for rescheduling, ensuring that the most impactful changes are addressed first.

Experimental Setup

The researchers evaluated ADATSA using a physical Kubernetes (v1.16.2) cluster consisting of 53 servers. The workload was designed to mimic real-world diversity, featuring 100 applications across five resource profiles: CPU-intensive, memory-intensive, disk-intensive, bandwidth-intensive, and balanced. Performance was benchmarked using Apache JMeter to simulate high-concurrency HTTP traffic, allowing for the measurement of response times and throughput under stress.

Results and Discussion

ADATSA was compared against the native Kubernetes scheduler, a PSO-based scheduler (PSOS), and a standard LA model (LAEAS).

  • Resource Imbalance Degree (RID): ADATSA outperformed all competitors, reducing resource imbalance by 19.56% to 23.24%.
  • Convergence and Scalability: As the number of tasks scaled (up to 2,500+), ADATSA reached a balanced state significantly faster than other models, demonstrating its suitability for large-scale deployments.
  • Quality of Service (QoS): The algorithm improved average response times by 2.05% to 4.72% and increased throughput by 4.04% to 6.37%. Crucially, ADATSA provided a much steadier response time profile during rebalancing phases compared to the native Kubernetes scheduler.

Conclusion

ADATSA represents a significant advancement in container cloud scheduling by integrating Learning Automata with stability-aware reward mechanisms. By "sensing" the environment and adapting its internal logic, the algorithm maintains cluster health without the computational overhead of complex heuristics. Future research will focus on automating the tuning of reward-penalty parameters and exploring Deep Reinforcement Learning (DRL) to manage even more complex, heterogeneous resource environments.

Summary updated

BibTeX
@article{zhu2021self,
  author = {Lilu Zhu and Kai Huang and Yanfeng Hu and Xianqing Tai},
  title = {{A Self-Adapting Task Scheduling Algorithm for Container Cloud Using Learning Automata}},
  journal = {IEEE Access},
  year = {2021},
  doi = {10.1109/access.2021.3078773},
}
Figure