I am trying to profile page faults for my cuda program using below command
nsys profile -t cuda --force-overwrite true --cuda-um-cpu-page-faults=true --cuda-um-gpu-page-faults=true -o trace_report ./cuda_program
It shows the below error
Items {
Type: ErrorText
Value: "Wrong event order has been detected when adding events to the collection:\nnew event ={ StartNs=190185147777 StopNs=190185371995 GlobalId=347136872939520 Event={ CudaUvmGpuPageFaultEvent=[{ Address=140386738634752 NumberOfPageFaults=3 FaultAccessType=1 DeviceId=0 },] } Type=120 }\nlast event ={ StartNs=190526069698 StopNs=190526324124 GlobalId=347136872939520 Event={ CudaUvmGpuPageFaultEvent=[{ Address=140389102125056 NumberOfPageFaults=3 FaultAccessType=1 DeviceId=0 },] } Type=120 }"
}
@skottapalli can you take a look at this one?
@turbocandy - thank you for reporting the bug.
- Which version of nsys are you using?
- Which target platform and OS are you trying to profile?
- What is the output of
nvidia-smi
command on the target system?
- Could you share the .nsys-rep file?
- To allow me to debug, could you share the repro? Does it repro 100% of the time when profiling your cuda program?
The version of nsys is ‘NVIDIA Nsight Systems version 2021.5.2.53-28d0e6e’
OS version ‘Ubuntu 20.04.4’,
Nvidia GeForce 3090 driver 470.199.02,
CUDA version 11.4
nyc_0.nsys-rep (13.8 MB)
Yes the above issue arises every time profiling the CUDA program
Thank you for the info. The version is pretty old. We have fixed a few bugs in nsys and the underlying CUPTI w.r.t --cuda-um-cpu-page-faults feature. Please use the latest 2023.3 version from https://developer.download.nvidia.com/devtools/repos or Nsight Systems | NVIDIA Developer and let us know if you are still seeing the bug.