Research Topic
I/O performance in ML models.
Supervisors
- Professor Julian Kunkel (University of Göttingen)
Thesis Abstract
Understanding Behavior in Parallel I/O using Deep Learning.
My PhD research investigates how I/O performance can be understood and predicted for machine-learning workloads on HPC systems. Using image, video, and largelanguage-model workloads, I analyze how application configurations influence I/O behavior and overall performance. Performance data are collected using profiling and tracing tools such as Score-P and DFTracer, and machine-learning models are used to predict I/O characteristics. The broader goal is to support more systematic and efficient performance analysis of AI workloads on HPC systems.
Publications
- Zoya Masih, Radita Liem, and Julian Kunkel. “Factors Impacting I/O Time Proportion in AI Workloads”. In: Proceedings of the 34th International Symposium on High-Performance Parallel and Distributed Computing. 2025, pp. 1–5.
- Zoya Masih, Radita Liem, and Julian Kunkel. “Robust I/O Characterization of Machine Learning Workloads Across Performance Analysis Tools”. In: Proceedings of the 35th International Symposium on High-Performance Parallel and Distributed Computing. HPDC ’26. Association for Computing Machinery, 2026, pp. 766–773. isbn: 9798400726408. doi: 10.1145/3806645.3816123. url: doi.org/10.1145/3806645.3816123.
Conference Presentations
- ISC Conference 2024 (IODC Workshop): Trace Analysis of ML workflows
using Vampir - Efficient Data Storage workshop 2024: Performance Analysis of ML Models
Using Vampir - NHR Conference 2024: IO characterization for Screw Conveyor Simulations
Using LIGGGHTS - Parallel Tools Workshop 2024 : Performance Analysis of ML Models Using
Vampir - ISC Conference 2025 (IODC Workshop): Assessing ML I/O Behavior Using
IO500 - NHR Conference 2025: Factors Impacting I/O Time Proportion in AI
Workloads - ISC Conference 2026 (IODC Workshop): A synthesis of ongoing work across
ML workloads