Workshop: 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS
Authors: Gretel De la Peña and Hadi Sahin (Florida International University); Renata Frez (Rede Nacional de Ensino e Pesquisa, Rio de Janeiro); and Italo Brito, Vasilka Chergarova, Luis Marin, Jeronimo Bezerra, and Julio Ibarra (Florida International University)
Abstract: Research and Education (R&E) backbone networks support the transfer of petabyte-scale scientific data under stringent timing constraints. However, their operational monitoring remains predominantly reactive, relying on threshold-based alerts and visualization dashboards that provide. limited predictive capabilities and cannot anticipate congestion or anomalous events hours in advance. Time Series Foundation Models (TSFMs), large neural networks pre-trained on billions of diverse time series, offer a promising route to predictive monitoring, but their adaptation behavior on live R&E backbone telemetry remains poorly characterized. We present a systematic multi-task benchmark of five TSFMs (TimesFM, Moirai, Chronos, PatchTST, and TTM/Granite) across four adaptation strategies and four auto-labeled operational tasks (48-hour utilization trajectory, threshold exceedance detection, peak-timing estimation, and traffic asymmetry forecasting), using 26 months of SNMP telemetry from the AmLight backbone with no human annotation. Our results show that TSFM performance is task-structured: zero-shot models outperform classical baselines on trajectory and timing tasks, where broad pre-training captures diurnal patterns, but do not surpass classical methods on tasks that reward high autocorrelation exploitation. Standard fine-tuning causes catastrophic forgetting in some architectures. Linear-probe fine-tuning avoids forgetting but introduces a subtler failure: auto-labels designed for one task do not transfer to others, so performance improves only where the label aligns with the target. This label-task mismatch is the central failure mode we identify and characterize. The unifying solution is a frozen backbone with a small, task-specific MLP head trained on the correct auto-label for each objective. With at most 500 auto-labeled windows, this approach surpasses all fine-tuned models on exceedance detection and matches fine-tuning on peak-timing estimation. These results establish that pre-trained TSFM backbones already encode operationally relevant network representations, and that a frozen backbone with a lightweight task head can be a viable foundation for predictive operations on R&E networks.
Back to 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS Archive Listing
Back to Full Workshop Archive Listing