Safe PD Capacity Forecasting with Time-Series Foundation Models and Calibrated Uncertainty for Heterogeneous GPU Clusters

Abstract

Heterogeneous GPU clusters need capacity forecasts that are accurate enough for planning and conservative enough for production-domain (PD) admission control. This paper evaluates safe PD capacity forecasting on Alibaba cluster-trace-gpu-v2023 using the released node and pod CSV files. We reconstructed an hourly time series from 1,523 nodes and 8,152 pods by combining scheduled pods, pending pods, GPU-sharing requests, CPU demand, memory demand, Quality-of-Service labels, pod phases, creation times, deletion times, and scheduled times. Because the released pod sample peaks at 59.49 effective GPUs while the physical cluster contains 6,212 GPUs, we define risk over explicit reserved PD slices rather than claiming whole-cluster exhaustion. The main risk setting is a 45-GPU PD quota, with additional quotas of 40, 50, and 55 GPUs. Six forecasting models were evaluated for 24-hour-ahead PD demand: persistence, ARIMA, Prophet-style additive regression, XGBoost, LSTM, and a compact multi-output time-series foundation-model proxy called TSFM-lite. Split conformal prediction calibrated 90% uncertainty intervals on the validation segment. The measured results show that ARIMA achieves the lowest test MAE at 5.82 effective GPUs, persistence obtains 6.08, and TSFM-lite obtains 6.93 while achieving 92.57% interval coverage. For safe risk detection at the 45-GPU quota, TSFM-lite catches 108 of 109 violation hours with one missed violation, while ARIMA catches all 109 but flags every test hour. The study demonstrates that calibrated intervals, not point forecasts alone, are the central mechanism for safe GPU capacity decisions.