DARO

This project investigates next-generation orchestration mechanisms for the emerging cloud–edge–IoT compute continuum, where applications must be deployed across highly heterogeneous, distributed, and resource-constrained infrastructures. Traditional cloud-native orchestration approaches, such as Kubernetes-based centralized scheduling, face challenges in supporting autonomous decision-making, dynamic resource environments, and non-container-native edge and IoT devices. The project aims to develop intelligent, adaptive, and decentralized orchestration solutions that manage complex distributed applications across the entire computing continuum.

A first outcome of this work is the Distributed Adaptive Cloud Continuum Architecture (DACCA), a Kubernetes-native architecture that extends orchestration capabilities beyond traditional data centers. DACCA introduces unified resource abstractions, decentralized self-awareness, and adaptive, swarm-based operation to enable resilient, context-aware workload management. At its core, the architecture incorporates the Distributed and Adaptive Resource Optimization (DARO) framework, which applies multi-agent reinforcement learning (MARL) to enable autonomous workload scheduling.

DARO introduces a decentralized scheduling approach in which distributed agents collaboratively optimize task placement using local observations and learned policies. Formulated as a decentralized partially observable Markov decision process (Dec-POMDP), the framework employs cooperative MARL techniques, including QMIX-based value decomposition, to train agents that can dynamically adapt to changing resource availability and workload characteristics. Experimental evaluations demonstrate the ability of the learned scheduling policies to achieve efficient and balanced resource utilization in dynamic cloud–edge-IoT environments.

To support the development, training, and evaluation of intelligent schedulers, the project also developed KWEST, a high-fidelity Kubernetes workload simulation framework. KWEST enables reproducible experimentation through synthetic workload generation, real-world workload trace replay, and flexible configuration of cluster and resource conditions. By providing realistic modeling of workload execution and scheduling behavior, the simulator facilitates systematic analysis and comparison of orchestration strategies without requiring costly physical deployments. Together, these contributions advance the vision of autonomous, AI-driven orchestration for future distributed computing infrastructures.