Where is the AR code?

[ConnollyLeon]

I have read the report about tutel, and I am curious about the implementation of AlltoAll.
I notice that tutel accelerate AlltoAll using CPU-GPU binding and adaptive routing tuning. I could find the CPU-GPU binding code in system_init.py, but fail to find code corresponding to adaptive routing tuning.

Could you please help me to locate the code?

[abuccts]

Greetings! Adaptive Routing (AR) allows Azure NDv4 VM to automatically detect and avoid network congestion by dynamically selecting more optimal network paths in the InfiniBand networking. It requires cluster level configurations including InfiniBand switch etc. and cannot be controlled by application only. You can leverage its tuning if using Azure HPC image (which contains optimized environment and packages) with multiple NDv4 VMs. Please refer to this article or Mellanox AR documents for more information about adaptive routing.