Dynamic prediction of resource consumption in virtualized environments

Abstract

Today’s cloud environments host various services and applications. In such an environment, the consumption of resources is growing exponentially which makes their management a challenging task especially in a dynamic cloud environment (e.g., resource availability, traffic volatility) and under different Quality of Service (QoS) requirements that are specified in Service-Level- Agreements (SLAs) for various applications (e.g., highly latency-sensitive applications). These applications designed as chains of microservices are deployed in a dynamic cloud environment. In this thesis, we focus on adapting resources for service function chains (SFCs) which consist of a set of Virtual Network functions (VNFs). High traffic load variations of applications lead to uncertainty in resource utilization. Consequently, the need for efficient data-driven mechanisms for automatic resource management becomes paramount. These mechanisms enable distributed systems to anticipate and efficiently respond to resource consumption fluctuations. They rely on accurate prediction techniques to satisfy the resource needs, to meet the QoS requirements for cloud applications and service infrastructures. These mechanisms are requiered to handle with the heterogeneity and the dynamic of resource demands. In this context, a mono-prediction model based on only historical consumption is inefficient for a volatile traffic workload. In this thesis, we propose a new prediction technique for dynamic workload variation. This technique allows the choice from various prediction techniques, the efficient one to trigger. Our solution applies a multi-task selector based on a meta-learning strategy MT-MLS. The MT-MLS introduces a new concept by analyzing similarities of resource consumption between various VNFs of an SFC. Additionally, an attention mechanism is employed to assign a weight to each Virtual Network Function (VNF) through similarity analysis. The MT-MLS is designed as a multi-task classifier, concurrently and independently assigning the most suitable predictor for forecasting multidimensional resource consumption for each Virtual Network Function (VNF) in a given workload. The best predictor is selected among a set of predictor models supported by our solution. Among the predictor models supported by our solution, the solution comprises an efficient Graph Neural Network (GNN) Model that uses SFC topological features demonstrating its performance under high-load traffic. The solution also includes an Long Short Term Memory (LSTM) model, a Convolutional Neural Network (CNN) model, and a hybrid model. These base predictors are used to generate metadata to train the MT-MLS. The performances of each model were compared with MT-MLS performances. The analysis of the various experimental results shows clearly that the proposed solution outperforms the other models.

Date	30 Jul 2024
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Nadjia Kara (Supervisor)