Performance Engineering in ML Workloads Using Score-P and Vampir

Content

Performance analysis is essential for understanding and optimizing the efficiency of ML/AI workloads, where computational and I/O bottlenecks can significantly impact training time, scalability, and resource utilization. By systematically identifying hotspots and inefficiencies, researchers and engineers can make informed decisions to improve both performance and energy efficiency. In this workshop, we introduce Score-P and Vampir, two powerful and complementary tools for performance measurement and visualization. Score-P provides detailed profiling and tracing of program behavior, while Vampir enables intuitive exploration of these traces through interactive visual analysis, helping users uncover complex performance patterns in ML workflows. The workshop consists of an introductory talk covering fundamental concepts, followed by a hands-on session where participants will instrument an example ML workload, collect performance data, and analyze the results using Vampir.

Requirements

• Basic Linux command-line skills
• Foundational understanding of ML workloads and performance concepts
• Basic understanding of Python codes

Learning goal

• Having a good understanding of performance analysis and its impact on AI workloads
• An understanding of the difference between I/O and computation bottlenecks
• Getting familiar with Profiling tool Score-P and visualization tool Vampir
• Getting knowledge on how to use the tools in analyzing the performance of a code

Skills

Trainer

• Zoya Masih

Next appointment