Nvprof

Nvprof is a command-line light-weight GUI-less profiler available for Linux, Windows, and Mac OS. It is capable of providing a textual report :

• Summary of GPU and CPU activity
• Trace of GPU and CPU activity
• Event collection

Nvprof also features a headless profile collection with the help of the Nvidia Visual Profiler:

• First use Nvprof on headless node to collect data
• Then visualize timeline with Visual Profiler

Quickstart guide

Environment modules

Before you start profiling with NVPROF, the appropriate module needs to be loaded.

NVPROF is part of the CUDA package, so run module avail cuda to see what versions are currently available with the compiler and MPImodules you have loaded. For a comprehensive list of Cuda modules, run module -r spider '.*cuda.*'. At the time this was written these were:

• cuda/10.0.130
• cuda/10.0
• cuda/9.0.176
• cuda/9.0
• cuda/8.0.44
• cuda/8.0

Use module load cuda/version to choose a version. For example, to load the CUDA compiler version 10.0, do:

[name@server ~]$ module load cuda/10.0