Horizontal and Vertical Scaling of Container-Based Applications Using Reinforcement Learning

Rossi, Fabiana; Nardelli, Matteo; Cardellini, Valeria

doi:10.1109/cloud.2019.00061

Cited by 116 publications

(59 citation statements)

References 21 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Elastic container deployment in cloud computing. The problem of elastic container deployment in Cloud computing has been studied intensively at different resource levels: container deployment, 16‐19 cluster deployment 20‐22 or both 23,24 . These approaches use horizontal methods, 17,21‐23 vertical methods 18,24 or hybrid approaches 16,19,20 depending on the elasticity dimensions.…”

Section: Related Workmentioning

confidence: 99%

“…The problem of elastic container deployment in Cloud computing has been studied intensively at different resource levels: container deployment, 16‐19 cluster deployment 20‐22 or both 23,24 . These approaches use horizontal methods, 17,21‐23 vertical methods 18,24 or hybrid approaches 16,19,20 depending on the elasticity dimensions. However these solutions lack the ability to include spatial aspects in their adaptation processes, which is essential to reduce network latency—a key performance factor for global web applications.…”

Section: Related Workmentioning

confidence: 99%

“…Such manual adaptation decisions are naive and inefficient as they can lead to costly, over‐provisioning issues (excessive deployment of clusters), performance issues (improper cluster placement) and/or adaptation issues, for example, performance degradation during adaptation and/or unnecessary cluster relocation. Much of the prior work on elastic container deployment problems in the Cloud does not consider spatial aspects of workload variations as they focus on local techniques within a single data center to handle fluctuations in workload volumes through localized auto‐scaling of containers, 16‐19 clusters, 20‐22 or both 23,24 . Other work proposes techniques to handle geo‐workload variations, 11,25 however, they do not use containers and hence do not benefit from the benefits of container‐based solutions, and they do not maintain application performance during adaptation.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Elastic deployment of container clusters across geographically distributed cloud data centers for web applications

Aldwyan

Sinnott

Jayaputera

2021

Concurrency and Computation

View full text Add to dashboard Cite

Containers such as Docker provide a lightweight virtualization technology. They have gained popularity in developing, deploying and managing applications in and across Cloud platforms. Container management and orchestration platforms such as Kubernetes run application containers in virtual clusters that abstract the overheads in managing the underlying infrastructures to simplify the deployment of container solutions. These platforms are well suited for modern web applications that can give rise to geographic fluctuations in use based on the location of users. Such fluctuations often require dynamic global deployment solutions. A key issue is to decide how to adapt the number and placement of clusters to maintain performance, whilst incurring minimum operating and adaptation costs. Manual decisions are naive and can give rise to: over‐provisioning and hence cost issues; improper placement and performance issues, and/or unnecessary relocations resulting in adaptation issues. Elastic deployment solutions are essential to support automated and intelligent adaptation of container clusters in geographically distributed Clouds. In this article, we propose an approach that continuously makes elastic deployment plans aimed at optimizing cost and performance, even during adaptation processes, to meet service level objectives (SLOs) at lower costs. Meta‐heuristics are used for cluster placement and adjustment. We conduct experiments on the Australia‐wide National eResearch Collaboration Tools and Resources Research Cloud using Docker and Kubernetes. Results show that with only a 0.5 ms sacrifice in SLO for the 95th percentile of response times we are able to achieve up to 44.44% improvement (reduction) in cost compared to a naive over‐provisioning deployment approach.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Elastic deployment of container clusters across geographically distributed cloud data centers for web applications

Aldwyan

Sinnott

Jayaputera

2021

Concurrency and Computation

View full text Add to dashboard Cite

show abstract

“…Garí et al [21] conducted a survey of previous RL solutions for scaling and scheduling problems in the cloud. Rossi et al [22] compared the Q-learning, the Dyna-Q, and a full backup model-based Q-learning to autoscale Docker Swarm containers horizontally and vertically. They measured the transition rate between different CPU utilization values to estimate the model, however this method is difficult to scale as it needs to store the number of transitions between every state.…”

Section: Related Workmentioning

confidence: 99%

Scaling UPF Instances in 5G/6G Core With Deep Reinforcement Learning

Nguyễn

Rotter³

2021

IEEE Access

View full text Add to dashboard Cite

In the 5G core and the upcoming 6G core, the User Plane Function (UPF) is responsible for the transportation of data from and to subscribers in Protocol Data Unit (PDU) sessions. The UPF is generally implemented in software and packed into either a virtual machine or container that can be launched as a UPF instance with a specific resource requirement in a cluster. To save resource consumption needed for UPF instances, the number of initiated UPF instances should depend on the number of PDU sessions required by customers, which is often controlled by a scaling algorithm. In this paper, we investigate the application of Deep Reinforcement Learning (DRL) for scaling UPF instances that are packed in the containers of the Kubernetes container-orchestration framework. We propose an approach with the formulation of a threshold-based reward function and adapt the proximal policy optimization (PPO) algorithm. Also, we apply a support vector machine (SVM) classifier to cope with a problem when the agent suggests an unwanted action due to the stochastic policy. Extensive numerical results show that our approach outperforms Kubernetes's built-in Horizontal Pod Autoscaler (HPA). DRL could save 2.7-3.8% of the average number of Pods, while SVM could achieve 0.7-4.5% saving compared to HPA.

show abstract

“…In contrast, such an approach is widely used, including Amazon's EC2, a virtual machine-based cloud platform. However, for applications that are constantly changing their requirements for CPU, memory, and other resources, their performance and resource utilization decrease significantly [5][6][7].…”

Section: Introductionmentioning

confidence: 99%

A Fine-Grained Horizontal Scaling Method for Container-Based Cloud

Jiang

2021

Scientific Programming

View full text Add to dashboard Cite

The container scaling mechanism, or elastic scaling, means the cluster can be dynamically adjusted based on the workload. As a typical container orchestration tool in cloud computing, Horizontal Pod Autoscaler (HPA) automatically adjusts the number of pods in a replication controller, deployment, replication set, or stateful set based on observed CPU utilization. There are several concerns with the current HPA technology. The first concern is that it can easily lead to untimely scaling and insufficient scaling for burst traffic. The second is that the antijitter mechanism of HPA may cause an inadequate number of onetime scale-outs and, thus, the inability to satisfy subsequent service requests. The third concern is that the fixed data sampling time means that the time interval for data reporting is the same for average and high loads, leading to untimely and insufficient scaling at high load times. In this study, we propose a Double Threshold Horizontal Pod Autoscaler (DHPA) algorithm, which fine-grained divides the scale of events into three categories: scale-out, no scale, and scale-in. And then, on the scaling strength, we also employ two thresholds that are further subdivided into no scaling (antijitter), regular scaling, and fast scaling for each of the three cases. The DHPA algorithm determines the scaling strategy using the average of the growth rates of CPU utilization, and thus, different scheduling policies are adopted. We compare the DHPA with the HPA algorithm under different loads, including low, medium, and high. The experiments show that the DHPA algorithm has better antijitter and antiload characteristics in container increase and reduction while ensuring service and cluster security.

show abstract

Horizontal and Vertical Scaling of Container-Based Applications Using Reinforcement Learning

Cited by 116 publications

References 21 publications

Elastic deployment of container clusters across geographically distributed cloud data centers for web applications

Elastic deployment of container clusters across geographically distributed cloud data centers for web applications

Scaling UPF Instances in 5G/6G Core With Deep Reinforcement Learning

A Fine-Grained Horizontal Scaling Method for Container-Based Cloud

Contact Info

Product

Resources

About