I want to test the number of bytes transfered by NVLink on V100, GPUs are connected by NVLink. But when I use metric
nvltx__bytes, ncu say fail to find metric. And when I use
Nvlink_Tables section, ncu say no metrics to collect found in sections.
How to solve above problems? Thanks.
Maybe I didn’t use
nvltx__bytes correctly, if I add
.sum suffix, it’s ok. But I still cannot make
Nvlink_Tables section work.
Note that the NVLink sections are not part of the default set.
So to get Nvlink_Tables section you can either use the ncu option –set full or –section Nvlink_Tables.
Nvlink_Tables section, but not work, and ncu dose not print any metric.
I think displaying tables from the Nvlink_Tables section from the ncu command line is not supported. You can try using “–section Nvlink --section Nvlink_Tables -o prof.ncu-rep” and load prof.ncu-rep in the ncu-ui.
Also what is the version of Nsight Compute you are using? You can use: ncu --version
The version of ncu I used is
Version 2022.1.1.0 (build 30914944) (public-release). The command I used is
sudo /usr/local/cuda-11/bin/ncu --section Nvlink --section Nvlink_Tables -o prof.ncu-rep ./a.out
prof.ncu-rep in ncu-ui, below is some screenshot:
As you can see, no nvlink table.
This seems to be a bug in ncu version 2022.1.1. Can you please try using the latest ncu version? You can download this from Gameworks Download Center | NVIDIA Developer.
Also note that to view the Nvlink topology you will also have to use the option –section Nvlink_Topology.
Or you can also use –section regex:Nvlink.* to select all Nvlink sections.