Software Diagnostics with NVIDIA Nsight Compute
NVIDIA Nsight Compute is an interactive kernel profiler for CUDA applications. It provides detailed performance metrics and API debugging via a user interface and command line tool.
Methods of Installation:
You can Load NVIDIA Nsight Compute with any version of CUDA as a module (recommended method)
You can install on your local machine and run on a remote server.
Load NVIDIA Nsight Compute as a Module
Nvidia Nsight Compute is available in all versions of CUDA available on PSMN servers.
Load the module for the CUDA version of your choice
module use /applis/PSMN/debian11/E5/modules/all
module load CUDA/12.4.0
Navigate to the Nsight Compute folder
cd /applis/PSMN/debian11/E5/software/CUDA/12.4.0/nsight-compute-2024.1.0
To run nsight compute command line interface from any machine
./ncu
To run nsight compute graphical user interface from a visual machine
./ncu-ui
Installing NVIDIA Nsight Compute on your local machine
You may download the version of Nsight Compute for your system at: https://developer.nvidia.com/tools-overview/nsight-compute/get-started
Note
You will likely be required to create an account with NVIDIA in order to download Nsight Compute.
Nsight Compute permits access to remote servers via proxy jump. Therefor, you may install Nsight Compute on your workstation and run it on a remote machine.
In order to use this method be sure the .ssh/config
on your
local machine is configured for a proxyjump.
Hostname r730gpu01
User $mylogin
Proxyjump allo-psmn # from inside ENS network
(mylogin
corresponding to your login at PSMN)
Navigate to the Nsight Compute folder and launch the GUI.
user@localmachine:~/NVIDIA-Nsight-Compute-2024.3$ ./ncu-ui &
Configure a New Remote Connection:
- Choose Start Activity
Connection: Click
+
- Target Platform
Choose Linux (x86_64)
- Add Remote Connection
Authentication Mode : Private Key
Enable TCP Tunnelling: Select
IP/Hostname: r730gpu01
ADD
Tip
Be sure to use the same hostname that is in your .ssh/config
file!