Using ncu to profile pagefault in unified memory

Question

is there any option to profile unified virtual memory CUDA application with Nsight Compute (NCU)? For example, I want to know the time spending on handling page fault and migration.

you would typically use nsight systems for this not nsight compute — Robert Crovella, Jul 01 '21 at 23:55
Hi, when I use the command like `nsys profile --stats=true --cuda-um-gpu-page-faults=true --cuda-um-cpu-page-faults=true ./myapplication` it does not showup any profiling statistics about those page faulting in the output, there any step I was missing? @RobertCrovella — Daniel, Jul 05 '21 at 22:38

score 0 · Accepted Answer · answered Jul 12 '21 at 01:08

Finally, I figure out the solution by myself. Just need to specify --export=json to output the profiling result into json file to get the detailed metrics of page fault. The overall profiling command looks like this.

nsys profile \
     --force-overwrite=true \
     --cuda-um-gpu-page-faults=true \
     --cuda-um-cpu-page-faults=true \
     --export=json \
     ./yourapplication

Using ncu to profile pagefault in unified memory

1 Answers1