What's the difference when starting tritonserver with mpirun --allow-run-as-root -n 1 /opt/tritonserver/bin/tritonserver vs. /opt/tritonserver/bin/tritonserver directly?

[so2bin]

Description
I am observing a difference in the behavior of TritonServer when starting it with mpirun compared to starting it directly. Specifically, when I use mpirun --allow-run-as-root -n 1 /opt/tritonserver/bin/tritonserver, the server runs and inference normally, but when I start it directly with /opt/tritonserver/bin/tritonserver, I notice higher CPU usage and slower inference speeds.

• The following is the normally started tritonserver's CPU usage from top and process information from ps auxfww:

  • The CPU usage information:
    

  • The processes information:
    

• The following is the abnormally started tritonserver's CPU and GPU informations:

  • The CPU usage information:
    

  • The nvidia-smi information:
    

  • The blocked start logs result from the high CPU usage:
    

• The following is the model files:
  

• Notes:

1. Whole the configurations of these two runnings are the same, only the start command changes.
2. All the models are running with Triton python backend.

I am wondering if there are any specific configuration that mpirun --allow-run-as-root -n 1 sets that might be influencing the behavior of TritonServer.
Thank you for your help in clarifying this matter.

Triton Information

To Reproduce

Expected behavior