The transition from petascale to exascale systems requires HPC research to investigate I/O performance more than ever before to assist application developers as well as system owners to achieve the best possible performance. In addition to the behaviours of different applications the congestion effects, global I/O weather and system noise come into play. Recent results such as [6] and [5] by Isakov et al. demonstrate that Machine Learning based modeling is a promising tool to cope with this complexity. However, the approach by Isakov et al. requires large amounts of training data which is the reason why Dmytro Povaliaev proposes a transfer learning approach in his master thesis [15]. Motivated by the mentioned previous work, I show a novel deep dive analysis workflow which allows to analyse the predictions quality of a model on the level of individual applications or even lower. Additionally, the integration of explainable AI algorithms enables the practitioner of this workflow to gain detailed insights into the I/O patterns the model has learned. Using this workflow I demonstrate that my model is able to predict the time widely used HPC applications spent on I/O with an accuracy that is considered to be usable in practice by domain experts and system owners. Finally, my workflow allows to isolate insufficient predictions and improve them by further fine tuning the model.

Masterarbeit, RWTH Aachen University, 2024; Aachen : RWTH Aachen University 1 Online-Ressource: Illustrationen (2024). = Masterarbeit, RWTH Aachen University, 2024

Published by RWTH Aachen University, Aachen