Unable to get metrics with prefix 'pmsampling:' in Jetson orin nx

[syrenisk]

I was trying to get the unified memory throughput while executing Gemm-ops based on PyTorch.
Following the manual, I executing following instruction:

ncu --query-metrics-collection pmsampling

and then I got a list of options available on Orin.
I chosn some related to the unified memory by key words ‘soc’:

mcc__dram_throughput_op_read                                                Throughput      %               soc_mcc             MCC read throughput                                                                   
mcc__dram_throughput_op_read_internal_activity                              Throughput      %               soc_mcc             MCC read throughput, internal activity                                                
mcc__dram_throughput_op_write                                               Throughput      %               soc_mcc             MCC write throughput                                                                  
mcc__dram_throughput_op_write_internal_activity                             Throughput      %               soc_mcc             MCC write throughput, internal activity                                               
mcc__dram_throughput_srcnode_cpu_op_read                                    Throughput      %               soc_mcc             MCC read throughput from CPU                                                          
mcc__dram_throughput_srcnode_cpu_op_read_internal_activity                  Throughput      %               soc_mcc             MCC read throughput from CPU, internal activity                                       
mcc__dram_throughput_srcnode_cpu_op_write                                   Throughput      %               soc_mcc             MCC write throughput from CPU                                                         
mcc__dram_throughput_srcnode_cpu_op_write_internal_activity                 Throughput      %               soc_mcc             MCC write throughput from CPU, internal activity                                      
mcc__dram_throughput_srcnode_dbb_op_read                                    Throughput      %               soc_mcc             MCC read throughput from DBB                                                          
mcc__dram_throughput_srcnode_dbb_op_read_internal_activity                  Throughput      %               soc_mcc             MCC read throughput from DBB, internal activity                                       
mcc__dram_throughput_srcnode_dbb_op_write                                   Throughput      %               soc_mcc             MCC write throughput from DBB                                                         
mcc__dram_throughput_srcnode_dbb_op_write_internal_activity                 Throughput      %               soc_mcc             MCC write throughput from DBB, internal activity                                      
mcc__dram_throughput_srcnode_gpu_op_read                                    Throughput      %               soc_mcc             MCC read throughput from GPU                                                          
mcc__dram_throughput_srcnode_gpu_op_read_internal_activity                  Throughput      %               soc_mcc             MCC read throughput from GPU, internal activity                                       
mcc__dram_throughput_srcnode_gpu_op_write                                   Throughput      %               soc_mcc             MCC write throughput from GPU                                                         
mcc__dram_throughput_srcnode_gpu_op_write_internal_activity                 Throughput      %               soc_mcc             MCC write throughput from GPU, internal activity

I executed some of them, just as following instructions:

sudo /usr/local/cuda-12.6/bin/ncu \
      --nvtx \
      --set full\
      --metrics pmsampling:mcc__dram_throughput_op_read.avg.pct_of_peak_sustained_elapsed \
      --force-overwrite \
      -o "$ncu_report_file" \
      $CONDA_PYTHON_PATH $script_name

as well as options:

mcc__dram_throughput_op_read_internal_activity                              Throughput      %               soc_mcc             MCC read throughput, internal activity                                                
mcc__dram_throughput_op_write                                               Throughput      %               soc_mcc             MCC write throughput                                                                  
mcc__dram_throughput_op_write_internal_activity                             Throughput      %               soc_mcc             MCC write throughput, internal activity                                               
mcc__dram_throughput_srcnode_cpu_op_read                                    Throughput      %               soc_mcc             MCC read throughput from CPU

then I got some ‘*.ncu-rep’ files. I downloaded these files and open it with Nsight Compute GUI on windows11.
But when I tried to check, there was no metrics, just like following figure:

I tried to rename options, like:

pmsampling:mcc__dram_throughput_op_read
pmsampling:mcc__dram_throughput_op_read.avg

but didn’t work.
I tried other options, for example:

pmsampling:l1tex__data_pipe_lsu_wavefronts.avg

then I got the metrics:

So how can I solve this problem?
Platform:
Jetson orin nx
cuda 12.6
ncu 2024.3.1.0

[TomNVIDIA]

Hello,

Thanks for visiting the NVIDIA Developer forums.

Your topic will be best served in the Jetson category, I have moved this post for better visibility.

Cheers,
Tom

[syrenisk]

thank you

[AastaLLL]

Hi,

Thanks for reporting this.

With your command, we also see “no data available” with the Nsight compute.
We need to check if further and update.

[syrenisk]

Thank you.
The following picture shows that metrics such as ‘dramc_…’ cannot be collected on shared-memory soc like Jetson Orin NX. So how can I get the likely metrics such as shared-memory Throughput, Read Bandwidth and Write Bandwidth?