I’m trying to profile some code running on a Tesla T4, and it looks like I can’t use nvvp like I used to on the Tesla V100. I’m trying to use nv-nsight-cu-cli now, and when I run it with --query-metrics I get 1000s of metrics that I can collect. Is there any documentation anywhere to maybe help me sort through and figure out which metrics I might actually want to collect?
Thanks!
Nsight Compute comes with a selection of ‘section’ files, which group selections of metrics into logical sections. For example, there is a section for very high-level metrics (Speed Of Light), one for memory-related metrics (Memory Workload Analysis) and so on. In addition, those sections are roughly ordered high- to low-level information. You get the pre-defined selection when just running nv-nsight-cu-cli, or you can select a sub-set using the --section option. Finally, these sections are just text files in the installation’s directory (you can use --list-sections to see their location), and you can inspect and modify them to see what metrics we think are relevant.