View source for Nsight-Compute-CLI-Sapelo2

[[Category:Sapelo2]][[Category:Software]][[Category:Other]]  

=== 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]]

The compute nodes that are connected to the GPU cards can access all the 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:
<pre class="gcommand">
module load CUDA/13.1.0
</pre>

* 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:
<pre class="gcommand">
module load CUDA/12.8.0
</pre>

* 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:
<pre class="gcommand">
module load CUDA/12.6.0
</pre>

* 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:
<pre class="gcommand">
module load CUDA/12.4.0
</pre>

* 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:
<pre class="gcommand">
module load CUDA/12.1.1
</pre>


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
<pre class="gcommand">
export TMPDIR=/scratch/$USER
</pre>

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 <code>--nodelist</code> option. For example, <code>--nodelist=ra8-3 </code>. 

'''Sample command to request an interactive job on a specific node:'''
<pre class="gcommand">
interact -p gpu_p --gres=gpu:H100:1 --nodelist=ra8-3
</pre>

'''Sample Slurm header line to request a specific node for a batch job:'''
<pre class="gcommand">
#SBATCH --nodelist=ra8-3
</pre>


=== Documentation ===

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

=== Installation ===

Downloaded from NVIDIA site.

=== System ===

64-bit Linux