Answers are grounded in this paper and cite their sources.
Cloud data centers face significant challenges from co-location interference, where virtual machines sharing physical hosts compete for CPU, memory, and I/O resources. Existing placement strategies commonly build on bin-packing heuristics and their multi-objective or learning-based extensions (Section 2.1), which optimize primarily at provisioning time and do not fully capture the dynamic, time-varying nature of workload interactions. We propose a Heterogeneous Temporal Graph Neural Network (HT-GNN) framework for interference-aware virtual machine (VM) placement and proactive migration scheduling. The framework models data center infrastructure as a heterogeneous graph comprising VM nodes, host nodes, runs-on edges, and co-location edges. We combine a bidirectional Graph Attention Network (GAT) with a Long Short-Term Memory (LSTM) encoder to jointly capture structural dependencies and temporal workload patterns. We train the model on empirically calibrated synthetic traces obtained from the Azure VM Noise dataset and the Alibaba GPU Cluster Trace (2023), enabling multi-task prediction of interference risk scores and Service Level Agreement (SLA) violation events. Our experiments show a 42.4% reduction in MAE and an AUC-ROC of 0.899, compared to 0.711 for a Gradient Boosting Trees (GBT) baseline and 0.763 for an LSTM-only baseline. Ablation experiments confirm that the LSTM temporal encoder and co-location edges are the most impactful components. The proactive migration policy achieves 6.8% higher host utilization than a random-placement baseline and 3.5% higher than a reactive threshold-based baseline, under matched exogenous workload conditions.
No takes yet. Share an insight, caveat, or question.
Pote et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: