Banca de QUALIFICAÇÃO: THIAGO PEREIRA DA SILVA

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : THIAGO PEREIRA DA SILVA
DATE: 20/07/2022
TIME: 14:00
LOCAL: Google Meet
TITLE:

An Ensemble Online Learning-based Approach for VNF Scaling in the Edge


KEY WORDS:

Auto-scaling, Edge Computing, Online Machine Learning, Ensemble Learning, Virtual Network Functions.


PAGES: 80
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUBÁREA: Sistemas de Computação
SPECIALTY: Teleinformática
SUMMARY:
Edge Computing (EC) platforms have recently been proposed to manage emergence applications with high computational load and low response time requirements. These platforms take advantage of the distributed nature of edge devices close to end-users and data sources, thus minimizing constraints such as bandwidth consumption, network congestion, response time, and operational costs imposed by cloud providers. There are different implementations of the concept of Edge Computing, such as Multi-Access Edge Computing (MEC). The European Telecommunications Standards Institute (ETSI) specified the ETSI MEC, a new ecosystem and value chain that offers application developers and content providers cloud-computing capabilities and an Information Technology (IT) service environment at the edge of the network. In order to provide more agility and flexibility for service provisioning while reducing deployment costs for infrastructure providers, technologies such as Network Functions Virtualization (NFV) are frequently used in production environments. NFV promotes the decoupling of hardware and network functions using virtualization technologies, enabling them to run in virtual machines or containers as software. Network functions or even higher layers functions are implemented as software entities called Virtual Network Functions (VNFs). Therefore, the main advantage of NFV is allowing multiple VNFs to run on just one server and scale them to consume the remaining free resources. The integration of Edge Computing and NFV paradigms, as proposed by ETSI MEC, enables the creation of an ecosystem for 5G applications. Such integration allows the creation of VNF chains, representing end-to-end services for end-users and their deployment on edge nodes. A Service Function Chaining (SFC) comprises a set of VNFs chained together in a given order, where each VNF can be running on a different edge node. The main challenges in this environment concern the dynamic provisioning and deprovisioning of distributed resources at the edge to run the VNFs and meet application requirements while optimizing the cost to the infrastructure provider. In this sense, scaling VNFs in this environment represents creating new containers or virtual machines and reallocating resources to them. However, manually performing the scaling actions is impractical, and auto-scaling approaches are required due to the variation in the workload and dynamic nature of the EC environment. This work presents a hybrid auto-scaling approach for the dynamic scaling of VNFs in the edge computing environment. Such an auto-scaling approach employs an online ensemble machine learning technique that consists of different base online machine learning models that predict the future workload. The architecture of such an auto-scaling approach follows the abstraction of the MAPE-K (Monitor-Analyze-Plan-Execute over a shared Knowledge) control loop to dynamically adjust the number of resources in response to workload changes. This approach is innovative because it proactively predicts the workload to anticipate scaling actions and behaves reactively when the prediction model does not meet the desired quality. In addition, our solution requires no prior knowledge of the behavior of the data, which makes it suitable for use in different contexts. We also have developed an algorithm to scale the VNF instances in the edge computing environment that uses a strategy to define how many resources to allocate or deallocate to a VNF instance during a scaling action. Finally, we evaluated the ensemble method and the proposed algorithm, comparing prediction performance and the amount of scaling actions and SLA violations.

COMMITTEE MEMBERS:
Presidente - 1213777 - THAIS VASCONCELOS BATISTA
Externo ao Programa - 2510306 - FREDERICO ARAUJO DA SILVA LOPES - UFRNExterna à Instituição - ATSLANDS ROCHA - UFC
Externa à Instituição - FLAVIA COIMBRA DELICATO - UFF
Externo à Instituição - PAULO DE FIGUEIREDO PIRES - UFF
Notícia cadastrada em: 23/06/2022 08:22
SIGAA | Superintendência de Tecnologia da Informação - (84) 3342 2210 | Copyright © 2006-2024 - UFRN - sigaa04-producao.info.ufrn.br.sigaa04-producao