Understanding why your GPU-accelerated is slow using NVIDIA Nsight Systems

This training is given in the context of the Scynergy 2026 event. It provides participants with a hands-on introduction to profiling using NVIDIA NSigh-Systems on the MeluXina supercomputer.

🎯 Objectives

By the end of this workshop, you should be able to:

• Profile your GPU jobs on MeluXina
• Interpret key NSight-Systems trace metrics and timelines
• Identify common bottlenecks in GPU accelerated codes/applications (IO, compute, memory, synchronization, communication)
• Apply simple optimizations and validate improvements

🪧 Agenda

Today's training is composed of:

• Connection to MeluXina via OpenOnDemand (~10 minutes)
• Introduction to NVIDIA NSight-Systems (~30 minutes)
• How to generate a trace with nsys-profile
• Looking at (already collected) traces of a MonAI training on Meluxina
  • How to navigate on the NSight-Systems GUI
  • How to use the nsys CLI to get some stats
• Hands-on: making a MonAI (PyTorch based) training faster (~60 minutes)
• Generate your own traces
• Modify the code to accelerate it

---

💻 Demo/Hands-on Mix

• Hands-on Part: Settings things up
• Demo and Discussion: ️Getting to know the tool and profiling of a slow [MonAI](https://project-monai.github.io/) training code
• Hands-on Part: Profile and optimize a distributed GPU-accelerated code

ℹ️ About this training

Authors: Marco Magliulo, Emmanuel Kieffer, and Tom Walter

This training has been developed by the Supercomputing Application Services group at LuxProvide in the context of the EPICURE project.