Nsight-Compute-CLI-Sapelo2

Category

Other

Program On

Sapelo2

Version

2023.1.1.0, 2024.1.0.0, 2024.3.0.0, 2025.1.0.0, 2025.4.0.0

Author / Distributor

NVIDIA

Description

From https://docs.nvidia.com/nsight-compute/NsightComputeCli/index.html: "NVIDIA Nsight Compute CLI (ncu) provides a non-interactive way to profile applications from the command line. "

Running Program

Also refer to Running Jobs on Sapelo2

Compute nodes equipped with GPU cards have access to all applications installed in /apps.

• Version 2025.4.0.0 is installed with CUDA v. 13.1.0 in /apps/eb/CUDA/13.1.0

To use it, please first load the module with:

module load CUDA/13.1.0

• Version 2025.1.0.0 is installed with CUDA v. 12.8.0 in /apps/eb/CUDA/12.8.0

To use it, please first load the module with:

module load CUDA/12.8.0

• Version 2024.3.0.0 is installed with CUDA v. 12.6.0 in /apps/eb/CUDA/12.6.0

To use it, please first load the module with:

module load CUDA/12.6.0

• Version 2024.1.0.0 is installed with CUDA v. 12.4.0 in /apps/eb/CUDA/12.4.0

To use it, please first load the module with:

module load CUDA/12.4.0

• Version 2023.1.1.0 is installed with CUDA v. 12.1.1 in /apps/eb/CUDA/12.1.1

To use it, please first load the module with:

module load CUDA/12.1.1

Please note:

1. Before running ncu to profile an application, please set the environment variable TMPDIR to point to a directory of yours. For example, use

export TMPDIR=/scratch/$USER

2. To profile applications on the H100 devices, please use the ncu in CUDA version 12.6.0 or higher. The older versions of ncu do not support the Hopper architecture.

3. Currently the ncu command only works on the following nodes:

A100: ra4-2

H100: ra8-3, ra7-2, ra8-4

More nodes will be added to the list above, as ncu is enabled on them (it requires a node reboot). Please check back, if of interest.

You can request a specific node with the Slurm --nodelist option. For example, --nodelist=ra8-3 .

Sample command to request an interactive job on a specific node:

interact -p gpu_p --gres=gpu:H100:1 --nodelist=ra8-3

Sample Slurm header line to request a specific node for a batch job:

#SBATCH --nodelist=ra8-3

Documentation

Please see https://docs.nvidia.com/nsight-compute/NsightComputeCli/index.html

Installation

Downloaded from NVIDIA site.

System

64-bit Linux