MODEL OF A SELF-ADAPTIVE DISTRIBUTED RESOURCE MANAGEMENT SYSTEM IN CLOUD COMPUTING

Abstract

This paper presents a decentralized approach to self-adaptive resource management in cloud infrastructures based on a network of interacting agents equipped with local predictive models and a consensus protocol. The proposed architecture integrates workload forecasting, local telemetry analysis, and P2P state synchronization mechanisms, forming a distributed MAPE (Monitor–Analyze–Plan–Execute) control loop that operates without a central controller. At the level of each node, an autonomous agent is implemented with the capability of collective decision-making for real-time service scaling. Unlike traditional centralized autoscaling mechanisms, the proposed system distributes control logic across cluster nodes, enabling local decision-making and coordinated global behavior through consensus. Each agent continuously analyzes multi-dimensional telemetry data, predicts workload trends using neural network-based models, and exchanges state information with neighboring agents. Consistency of scaling decisions is ensured by a consensus protocol while preserving fault tolerance under partial failures.

To evaluate the effectiveness of the proposed approach, a simulation environment was developed that reproduces complex workload patterns and failure scenarios of Kubernetes clusters, including load spikes, node failures, and network partitions. A comparative analysis between the decentralized architecture and a classical centralized autoscaling mechanism was conducted under identical conditions. Experimental results demonstrate that the proposed P2P-based system provides a significantly faster response to sudden workload changes, reduces CPU overload events, shortens recovery time after node failures, and maintains stable scaling behavior during network partitions. Although the P2P architecture introduces additional communication overhead, its impact remains negligible relative to overall cluster traffic. The results confirm that decentralized self-adaptive resource management is an effective alternative to centralized autoscaling in dynamic cloud-native environments.