Digital Signal Feature Extraction for Graph-Based Host Classification in VM Placement

Abstract

This study proposes a host classification methodology based on signal analysis and graph representation as a pre-placement stage for virtual machines (VM) in a cloud computing environment. The Bitbrains dataset is utilized as a source of time-series data representing CPU, memory, disk, and network utilization. Each parameter is modeled as a discrete signal and analyzed in both time and frequency domains. The analysis is conducted using a fixed observation window and frequency-domain transformation to capture workload characteristics across multiple resources. The Fourier transform results indicate the dominance of low-frequency components, suggesting gradual workload variations. Spectral energy is calculated and normalized to identify quantitative differences between host conditions. The results show that the overloaded class contributes 98.9% of the total spectral energy, while the underloaded and balanced classes contribute only 0.7% and 0.4%, respectively. The extracted features are then integrated using a four-node graph model that connects all resource dimensions into a single structure. The aggregated graph score is employed for dynamic percentile-based classification. From a total of 1,239 analyzed hosts, the proposed method classifies 421 hosts as overloaded, 409 as underloaded, and 409 as balanced. These findings demonstrate that spectral characteristics combined with graph integration provide a quantitatively structured and adaptive host segmentation mechanism, where the resulting classification can support VM placement decisions by identifying underloaded, balanced, and overloaded host conditions.