I’m fairly new in using the nsight systems.
I’m trying to profile my deepstream application. How do I figure out where the bottleneck is in this ?
I have attached the nsight report.
report3.nsys-rep (6.9 MB)
Okay, from looking at the diagnostic warnings, I can see that you had some permissions issues and did not collect everything that you requested. Specifically you didn’t get CPU backtraces because your paranoid level was too high. You can run
nsys status -e
to get information about your system and what functionality is allowed.
From looking at the analysis summary, I can see that your CPU utilization was pretty low and it looks like you spent a LOT of time in futex.
What version of Nsys were you using? (there was a weird error about the CTK version and the nsys version).
From looking at what data did get collected and opening the GPU rows, I can see that there are a lot of gaps in the GPU timeline.
I would strongly suggest you get the permissions set to see the CPU backtraces so you can see what functions are blocking you. See Installation Guide :: Nsight Systems Documentation (direct link to section) to see how to set the paranoid level.
You may also want to read Optimizing CUDA Memory Transfers with NVIDIA Nsight Systems | NVIDIA Technical Blog. I wrote it about memory transfers, but there is also a lot of good information there about navigating the tool.