Noisy-Neighbor-Aware VM Degradation Risk Modeling with Unsupervised Residual Fusion

Abstract

Cloud virtual machines can experience abrupt benchmark degradation when shared infrastructure is exposed to contention, maintenance activity, VM drift, storage-path effects, or unfavorable placement. This paper presents a noisy-neighbor-aware degradation-risk pipeline for benchmark traces without claiming host-level causal attribution. The study uses three public data sources: the official Azure VM Noise Dataset 2024, the Server Machine Dataset (SMD) from OmniAnomaly, and the Azure Trace for Packing 2020. The Azure experiment covers 7,037,220 raw measurements across 776 CSV partitions and evaluates a deterministic, per-partition balanced modeling table of 1,048,635 rows. Because Azure does not publish co-tenant identities or root-cause labels, Azure F1 scores are reported against weak observed-degradation labels derived from severe raw performance and runtime tails rather than injected synthetic labels. SMD provides an external label-based sanity check with 708,420 test points and official anomaly labels. The proposed score is renamed NN-Aware Residual Fusion (NN-RF): it combines signed performance residuals, runtime residuals, temporal deltas, and previous-day cell pressure through a non-negative validation-calibrated fusion model; PCA reconstruction is treated as a candidate feature rather than as a temporal autoencoder. On the held-out Azure period, NN-RF achieves F1 0.938, AUROC 0.998, and recall 0.932 under the weak degradation task, while Robust MAD remains a strong residual baseline with F1 0.932. On SMD, a windowed neural autoencoder gives the best held-out F1, 0.299, among the evaluated external baselines. A capacity-constrained replay of 50,000 Azure Packing Trace admissions shows that heat-aware first-fit keeps the same acceptance rate as trace-order first-fit while reducing mean placement risk by 7.0%. The results support a narrower conclusion: benchmark residuals can provide useful risk signals for placement, but they should not be interpreted as definitive noisy-neighbor root-cause labels.