Missing PCIe througput metric in nsys-report

Hello,

I am trying to profile a CUDA workload (needleman-wunsch from rodinia) using the following command on an A100-

EDIT: I have MIG mode enabled. The report is generated when running is workload on a MIG instance. Not sure if that affects profiling.

sudo nsys profile --trace=cuda,nvtx,osrt --cuda-event-trace=false --gpu-metrics-devices=1 \
--gpu-metrics-set=ga100 --force-overwrite true -o needle_profile ./needle 8192 10

I get the following nsys-report -

It’s missing the PCIe bandwidth metric, which I need. How do I fix this issue?

I am attaching the report here -
needle_profile.zip (311.5 KB)

Thanks.

@pkovalenko

PCIe metrics are not available when profiling MIG instances. You may want to profile the full chip GPU for that.

The PCIe metrics are not available when in MIG where the MIG partition is not the full GPU. PCIe is a shared unit. The performance counter are not virtualized per GPU instance or compute instance.

The two options are to either:

  1. run with MIG disabled
  2. run with MIG that is a full device