Classification of thread profiles for scaling application behavior.

•We classify thread profiles along different program runs carried out at different scale.•Classification is based on unsupervised machine learning, it does not require any a priori knowledge.•Aiming at exascale analysis, we fit per-thread-class extrapolation models of thread profile.•Validation for different application domains as: graph analytics, fluid dynamic and radio-astronomy.•Extrapolation models have relative errors of less than 10% for a real life radio-astronomy application.