nsys: Nvidia GPU and CPU profiler
Available
Puhti: 2022.1.3.3
Mahti: 2021.3.3.2
Usage
The nsys profiling tool collects and views profiling data from the command-line. It enables the collection of a timeline of CUDA-related activities on both CPU and GPU, including kernel execution, memory transfers, memory set and CUDA API calls and events or metrics for CUDA kernels. The tool is very useful in identifying the high-level bottlenecks, hotspots and for determining which kernels should be targeted for optimization and analysis with the Nsight Compute tool. Profiling results are displayed in the console after the profiling data is collected, and may also be saved for later viewing by nsys-ui tool.
To use nsys, one needs to first load the CUDA module:
To profile a CUDA code, one then adds the command nsys before the normal
command to execute the code. Running is otherwise similar to that of any other
CUDA job on Puhti or Mahti.
An example of usage and output of nsys:
$ nsys profile -t nvtx,cuda -o <results_file> --stats=true --force-overwrite true ./a.out
Collecting data...
Processing events...
Capturing symbol files...
Saving temporary "/tmp/cristian/6584503/nsys-report-b4eb-c068-9292-3b17.qdstrm" file to disk...
Creating final output files...
Processing [==============================================================100%]
Saved report file to "/tmp/cristian/6584503/nsys-report-b4eb-c068-9292-3b17.qdrep"
Exporting 4657 events:
Generating CUDA API Statistics...
CUDA API Statistics (nanoseconds)
Time(%) Total Time Calls Average Minimum Maximum Name
------- -------------- ---------- -------------- -------------- -------------- -------------------------------------------------------------
85.3 323223522 4 80805880.5 128957 322811927 cudaMalloc
13.6 51524634 1 51524634.0 51524634 51524634 cudaDeviceReset
....
Generating CUDA Kernel Statistics...
CUDA Kernel Statistics (nanoseconds)
Time(%) Total Time Instances Average Minimum Maximum Name
------- -------------- ---------- -------------- -------------- -------------- -------------------------------------------------------------
100.0 22912 1 22912.0 22912 22912 multiply_add_kn(float*, float const*, float const*, float const*, int)
Generating CUDA Memory Operation Statistics...
CUDA Memory Operation Statistics (nanoseconds)
Time(%) Total Time Operations Average Minimum Maximum Name
------- -------------- ---------- -------------- -------------- -------------- -------------------------------------------------------------
79.0 2022300 3 674100.0 663903 692095 [CUDA memcpy HtoD]
21.0 536223 1 536223.0 536223 536223 [CUDA memcpy DtoH]
CUDA Memory Operation Statistics (KiB)
Total Operations Average Minimum Maximum Name
------------------- -------------- ------------------- ----------------- ------------------- ------------------------------------------------
3906.250 1 3906.250 3906.250 3906.250 [CUDA memcpy DtoH]
11718.750 3 3906.250 3906.250 3906.250 [CUDA memcpy HtoD]
Generating Operating System Runtime API Statistics...
Operating System Runtime API Statistics (nanoseconds)
Time(%) Total Time Calls Average Minimum Maximum Name
------- -------------- ---------- -------------- -------------- -------------- -------------------------------------------------------------
67.0 343435124 29 11842590.5 23172 100249843 poll
22.6 115645051 1102 104941.1 1286 25309244 ioctl
5.5 28249766 4 7062441.5 3763 15288473 fread
....
nsys supports many useful running options. For more details please check the nvidia documentation.
The report above can also be viewed using the graphical interface. The results of the analysis are saved in the the specified file, <results_file>.qdrep and can be viewed directly on the CSC servers running nsys-ui or copied on local computers and viewed using a local installation of the nsight-systems.