Nsight Systems

Overview

  • Tutorial: 45 min

    Objectives:

    1. Understand the basics of Nsight Systems and how to use it for profiling applications.

Nsight Systems is a profiling tool for CUDA applications. It helps developers identify performance bottlenecks in their CUDA kernels and applications.

First, load the necessary modules.

1module load gcc/10.3.0
2module load cuda/12.9.0

Then compile the matrix multiplication application with NVCC

1cd /scratch/vp91/$USER/intro-to-profiling/code/nsys
2
3nvcc -O0 -g tiled_matmul_multikernel.cu -o tiled_matmul

  • -g enables debugging information.

  • -O0 disables optimizations to make debugging easier.

Nsight Systems profiler

Now run the program with Nsight Systems to check for performance bottlenecks

1nsys profile --trace=cuda,nvtx,osrt --stats=true ./tiled_matmul

or submit the job script

1cd /scratch/vp91/$USER/intro-to-profiling/job_scripts/nsys
2
3qsub 1_matmul.pbs

This will generate a report file with the extension .nsys-rep.

Viewing the report

1module load cuda/12.9.0
2
3nsys-ui report1.nsys-rep

On Gadi this will usualy result in an error due to missing display. To get around this, launch a Virtual Desktop session on Gadi and run the above command in the terminal available on the Virtual Desktop.

Once launched you can explore the performance data using the various views and analysis tools provided by Nsight Systems.

Nsys

Key Points

  1. Nsight Systems is a powerful tool for profiling CUDA applications.

  2. It provides insights into kernel execution times, memory transfers, and other performance metrics.

  3. Using Nsight Systems can help identify performance bottlenecks and optimize CUDA applications.