Skip to content

profiling.md

Latest commit

 

History

History
87 lines (60 loc) · 2.78 KB

profiling.md

File metadata and controls

87 lines (60 loc) · 2.78 KB

Kernel profiling with Nsight Compute

Nsight Compute is a tool for evaluating individual kernel performance on NVIDIA hardware.

Here's a screengrab of a sample profile opened in Nsight Compute:

NSYS Screenshot

Installation

Follow the instructions on the install page

At time of writing, those steps were:

  1. Download the .run installer for Linux systems.

  2. Modify the permissions in the installer and execute with:

    sudo base nsight-compute-linux-2025.2.0.12-33266684.run

  3. Build the binary:

    mojo build --debug-level=line-tables -o a.out my_file.mojo

  4. In your remote environment (the one connected to the NVIDIA GPU), run the following command:

    /usr/local/cuda/bin/ncu --set full -f -o <output_build_name> --import-source yes \
  --source-folder /home/ubuntu/modular /home/ubuntu/modular/<file_name.mojo>

You should see a file called output_build_name.ncu-rep in the directory. This will generate a built file with everything you need to profile.

Usage

Once installed, you can:

  • SCP the build into your local machine
  • Drag and drop to Nsight
  • Profile it

You can also do this process remotely:

  • SSH into the remote environment:

  • Open the Start Activity Dialog, select an SSH-capable Target Platform, and click the + button

  • Choose your authentication method:

    • Password: enter IP/hostname, username, password, port (default is 22), deployment directory, and optionally a connection name
    • Private Key: enter IP/hostname, username, private key path, passphrase, port, deployment directory, and optionally a connection name
    • If using Coder, the default username and password is ubuntu, and the default IP is coder.workspace_name.main

  • Select Target Application (per profiling session)

  • Select the remote connection in the Start Activity Dialog

  • Browse the remote file system over SSH

  • Select the compiled binary of the kernel you want to profile

  • You are responsible for compiling this binary outside Nsight Compute (see above commands to compile)

  • Launch Activity (what Nsight Compute does)

  • Copies required profiling tools (e.g., ncu, support files) to the deployment directory on the remote machine

  • Executes the selected binary on the remote device

  • Viewing

  • A report file (e.g., .ncu-rep) generated in the deployment directory on the remote device

  • This report is copied back to the host and opened in the UI

  • All actions (file transfers, tunneling, execution, report retrieval) are logged in the Progress Log inside Nsight Compute

You are now ready to profile any kernels!