Hi, I am currently writing a script to merge data from all 4 trace reports that nsight systems currently provides. There is a kernexec report, cuda api trace report, nvtx trace report, and gpu trace report. A lot of these have some overlapping information, but still provide relevant details. My question is that is there anything that nsight systems is not providing in terms of reports to generate a “full” trace. I saw that there are two tables in the sql file that are not being queried by any reports. One is “PROFILER_OVERHEAD” and the other is “CUPTI_ACTIVITY_KIND_SYNCHRONIZATION”. How relevant are these to generating a trace of every single function call, api call, etc. How should I be approaching this?
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| Nsight System outputs "CUDA trace data was not collected." and there is no result for cuda kernels | 3 | 1690 | September 25, 2023 | |
| Not getting NVTX events from child processes on Linux | 8 | 1203 | June 23, 2021 | |
| How get MPI Trace data on CLI | 2 | 782 | February 24, 2022 | |
| No CUDA events collected | 6 | 2354 | August 25, 2021 | |
| Nsight compute system trace issue | 4 | 371 | March 22, 2024 | |
| NVTX doesn't appear in timeline | 1 | 691 | December 19, 2019 | |
| nsys CUDA trace works for threads, but not for subprocesses | 3 | 2317 | May 13, 2019 | |
| Nsys Does not Track CUDA Api events | 5 | 1029 | December 22, 2022 | |
| Nsight Systems not creating qdrep file | 2 | 1422 | August 21, 2024 | |
| No NVTX events if other APIs (cuda, osrt) are enabled | 0 | 673 | April 23, 2019 |