Repository navigation
Consider other profilers to use #264
Description
It seems like @jagoosw previously used Nsight for profiling. Consider whether there are other, potentially Julie-built-in, alternatives to use for comparison.
It seems like @jagoosw previously used Nsight for profiling. Consider whether there are other, potentially Julie-built-in, alternatives to use for comparison.
Activity
MarionBWeinzierl commentedon Jul 17, 2025
In general, here is some material which gives you an overview of existing tools (although I don't know how many apply to Julia) and concepts of performance analysis:
Durham runs an, I think annual, Performance Analysis workshop series , but this year's was already in spring, so too late for you - but you can keep your eyes open for next year.
For Nsight in particular, its own website is probably the best place to start: https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html
Mikolaj-A-Kowalski commentedon Aug 1, 2025
We should try the native profiler of
CUDA.jlpackage.It should be simple to do. If we have a script prepared for profiling with NSight, we just need to run it without NSight and make sure we set
externaltofalse.EDIT:
One thing to be careful about is how to capture the results since they are printed to the screen when run in REPL. Naive inclusion in a script prints no output 😅
Mikolaj-A-Kowalski commentedon Dec 4, 2025
I will close this issue since it is more of an note for the profiling project and we have in the end settled on using NVidia tools.