Intent-Aware Multi-Domain Cloud Continuum Coordination through Hierarchical Deep Reinforcement Learning

Abstract

The unified paradigm of the Computing Continuum promises significant performance improvements for modern enterprise services. Yet, in practice, automated, self-adaptive infrastructure management remains an unresolved challenge. The competing objectives, often specified as the intents of various actors, such as service owners and infrastructure providers, intensify this difficulty. In fact, the concrete computing scenario is a combination of a manifold of isolated infrastructure instances, each controlled by separate providers in different domains (e.g., network and computing), operating under their own policies and strategies, with minimal interfaces to external entities. Identifying fair, dynamic, and efficient solutions within a landscape of conflicting general objectives and multiple domains with only partial observability is both highly complex and essential for a truly successful services deployment in the Continuum.To address this challenge, we introduce InCoord, a novel, modular coordination approach based on Hierarchical Deep Reinforcement Learning (DRL). Given a specified service-infrastructure pair, DRL-based Domain Agents (DAs) manage individual resource domains autonomously, adapting the underlying resources to fulfill the infrastructure-level intent. At a higher level, we design a lightweight coordinator that, by adjusting infrastructure-level intent conditions, helps meet a global-level intent while balancing potential conflicts among DAs. Each InCoord agent is first swiftly trained within a synthetic environment to ensure immediate readiness for deployment. We evaluate their performance in an emulation that provides a realistic, complex setting that captures the non-linear dynamics of real environments, such as sudden load changes, and requires the coordination of network and computing resource instances. The findings demonstrate that InCoord consistently minimizes intent violations to only 16%. Compared to a siloed environment without dynamic coordination, the infrastructure struggles to achieve the overall system objectives, exhibiting, on average, almost twice as many intent violations.