Nsight Compute#
Nsight Compute is a kernel profiler for CUDA applications, which can also be used for API debugging. It supports Volta architecture GPUs and newer (SM 70+).
On Bede, Nsight Compute is provided by a number of modules, with differing versions of ncu.
Remote GUI is not available on Bede, but profile data can be generated on Bede via the CLI for local use.
You should use a versions of ncu that is at least as new as the CUDA toolkit used to compile your application (if appropriate).
module load nsight-compute/2022.4.1
module load nsight-compute/2022.1.0
module load nsight-compute/2020.2.1
module load cuda/12.0.1 # provides ncu 2022.4.1
module load cuda/11.5.1 # provides ncu 2021.3.1
module load cuda/11.4.1 # provides ncu 2021.2.1
module load cuda/11.3.1 # provides ncu 2021.1.1
module load cuda/11.2.2 # provides ncu 2020.3.1
module load nvhpc/23.1 # provides ncu 2022.4.0
module load nvhpc/22.1 # provides ncu 2021.3.0
module load nvhpc/21.5 # provides ncu 2021.1.0
module load nsight-systems/2023.4.1
module load cuda/12.3.2 # provides ncu 2023.3.1
module load cuda/12.2.2 # provides ncu 2023.2.2
module load cuda/12.1.1 # provides ncu 2023.1.1
module load cuda/11.8.0 # provides ncu 2022.3.0
module load cuda/11.7.1 # provides ncu 2022.1.0
module load cuda/11.7.0 # provides ncu 2022.2.0
module load nvhpc/24.1 # provides ncu 2023.3.1
Consider compiling your CUDA application using nvcc with -lineinfo or --generate-line-info to generate line-level profile information.
A common use-case for using Nsight Compute on HPC systems is to capture all available profiling metrics for a run of a target application, storing the information to a file on disk. This file can then be interrogated on a local machine using the Nsight Compute GUI.
For example, the following command captures the full set of metrics for an application using using the command line tool ncu.
ncu -o profile --set full ./myapplication <arguments>
Capturing the full set of metrics can lead to very long run times, as each kernel is replayed many times.
Rather than capturing the full set of metrics, a subset may be captured using the --set, --section and --metrics flags as described in the Nsight Comptue Profile Command Line Options table.
The scope of the section being profiled can also be reduced using NVTX Filtering; or by targetting specific kernels using --kernel-id, --kernel-regex and/or --launch-skip see the CLI docs for more information).
Once the .ncu-rep file has been downloaded locally, it can be imported into local Nsight CUDA GUI ncu-ui via ncu-ui profile.ncu-rep or File > Open > profile.ncu-rep in the GUI.
Note
Older versions of Nsight Compute (CUDA < v11.0.194) provided nv-nsight-cu-cli nv-nsight-cu rather than ncu and ncu-ui respectively.
The generated report file used the .nsight-cuprof-report extension rather than .ncu-rep.
More Information#
-
Use the following Nsight report files to follow the tutorial.