Distributed LSTM-based cloud resource allocation in Network Function Virtualization Architectures

Modern digital infrastructures are increasingly expected to adapt in real time to highly dynamic workloads. In environments where network services are virtualised and deployed across geographically distributed cloud resources, one of the most challenging tasks is deciding how much computational capacity should be allocated before demand actually arrives. Traditional reactive approaches often struggle in this scenario. Resource reconfigurations require time, and delays in provisioning can translate into degraded service quality, inefficiencies, and increased operational costs. For this reason, predictive intelligence is becoming a key component of next-generation network management systems. A recent line of research explores how deep learning models can anticipate future resource requirements and enable proactive allocation strategies. Rather than waiting for traffic variations to occur, the infrastructure continuously forecasts future processing demands and prepares the necessary resources in advance.

One of the most effective techniques for time-series prediction is represented by Long Short-Term Memory (LSTM) neural networks. These models are particularly suitable for capturing long-term dependencies in temporal data and have demonstrated strong performance in forecasting tasks. However, applying LSTM models at scale introduces a significant challenge. A centralized predictor requires collecting monitoring data from all network locations, generating substantial communication overhead and raising concerns related to scalability and data management. To overcome these limitations, a distributed learning framework has been developed in which each virtualised network function maintains its own local predictor. Instead of transferring monitoring data to a central entity, the individual predictors collaborate by exchanging only selected neural network parameters. This strategy allows each node to preserve local information while still benefiting from knowledge emerging across the entire infrastructure. The result is a cooperative prediction process capable of improving forecasting accuracy without requiring massive data transfers.

The proposed architecture associates an intelligent agent with each virtualised network function. Every agent continuously monitors the processing capacity required by the service it manages and trains a dedicated LSTM-based forecasting model. At regular intervals, the agents exchange a subset of their neural network weights through a distributed consensus mechanism. Rather than creating a global centralised model, the system enables local predictors to align their knowledge while maintaining their autonomy progressively. This approach differs from traditional federated learning paradigms. Instead of relying on a central coordinator to aggregate updates and redistribute a global model, the learning process remains fully decentralised. Information flows directly among participating entities, reducing dependency on centralized orchestration mechanisms. The architecture, therefore, combines two important objectives that are often difficult to achieve simultaneously: high prediction quality and low communication overhead.

The methodology has been evaluated in a realistic Network Function Virtualization environment inspired by carrier-grade infrastructures. Multiple virtualised network functions are deployed across distributed network sites, each generating its own processing-demand time series. Different learning strategies are compared. The first relies on a fully centralised predictor with access to all available data. The second adopts completely independent local predictors without any cooperation. The third implements the proposed distributed learning mechanism. The results show that cooperative distributed prediction consistently outperforms isolated local forecasting. While the centralised solution still achieves the lowest prediction error in some scenarios, the distributed approach obtains comparable accuracy while dramatically reducing communication requirements. Additional experiments investigate different communication patterns among network functions. Even when information exchange is restricted to neighbouring nodes or entities belonging to the same infrastructure site, prediction performance remains highly competitive.

The significance of this work extends beyond resource allocation alone. It demonstrates how distributed artificial intelligence can become a practical tool for managing complex infrastructures characterised by geographical dispersion, privacy constraints, and communication limitations. As cloud-native architectures continue to evolve and network services become increasingly dynamic, the ability to combine local autonomy with collective intelligence will likely become a defining characteristic of future orchestration platforms. Rather than concentrating all intelligence in a central controller, the network itself becomes capable of learning, adapting, and coordinating. In this vision, prediction is no longer a standalone analytical task but an emergent property of a distributed ecosystem of intelligent agents working together toward a common objective. The result is a more scalable, resilient, and efficient approach to infrastructure management, where artificial intelligence is embedded directly into the operational fabric of the network itself.