Bridging the Heterogeneity Gap: Data-Efficient Transfer Learning for Cross-Platform HPC Performance Modeling

Performance models are only as useful as their ability to generalize — yet every new HPC platform demands weeks of data collection before a model can predict reliably. The core barrier is not data volume; it is structural heterogeneity. Hardware performance counters differ in name, count, and order across architectures, and statistical distributions shift even when feature names match. Existing transfer learning methods either assume identical feature spaces or require retraining from scratch — both options are operationally untenable at scale. Without a solution, performance modeling remains tethered to each platform it serves, unable to transfer knowledge forward as hardware evolves.
We introduce a few-shot, test-time adaptation framework that learns to bridge structurally disjoint feature spaces without manual alignment or model retraining. A learned input alignment layer automatically maps target features onto the source model's latent space, and a stacked residual model captures relationships that alignment alone cannot resolve. Evaluated across 800 experiments spanning 11 HPC and 4 ML datasets, our method matches fully supervised models trained on 100× more data using just 1–5% of target samples — cutting data collection time from days to seconds — and reduces job turnaround time by 71% in a real-world HPC scheduling deployment. When a performance model can learn from a handful of observations what a supervised baseline needs hundreds to know, cross-platform intelligence stops being a research result and becomes an operational reality.