We are pleased to share that ERA researchers have presented a new paper at the 2026 International Joint Conference on Neural Networks (IJCNN) at the workshop Data-Driven Decision-Making: Uncertainty and Reliable Decision-Making by Generative AI (DDM-GenAI): “Multimodal Runtime Fingerprints: A Generative Modeling Approach for Containerized Workload Analysis”, authored by Gennaro Mellone, Antonio Caccioppoli, Gaetano Volpe, Alberto Savarese, and Fabio Angeletti (Fides Consulting, Time Vision, LUISS).

The work introduces execution fingerprints — compact 10-dimensional vectors derived from eBPF profiling data — as a structured way to represent the runtime behavior of containerized applications. A key finding is that these fingerprints exhibit a strongly bimodal distribution that cuts across nominal load regime boundaries, a structural property not captured by standard workload labels.

Three generative models were compared, with a mixture-aware VAE+GMM approach proving the most effective at preserving this structure — achieving a synthetic silhouette score of 0.934 vs. 0.923 on real data. The work also establishes a methodological point: marginal statistics alone are insufficient to evaluate synthetic system telemetry; geometric and clustering metrics are essential.

These results have direct implications for the ERA project’s goals: accurate generative models of workload behavior can support scheduling decisions, capacity planning, and energy optimization in data centers without requiring live execution of target services — a concrete step toward smarter, more sustainable IT operations.