Allow all in/out communications to and from the compute subnet
Closed this issue · 0 comments
xpillons commented
compute nodes may have to communicate on ports which are already closed. This is the case for NCCL for which there is no way to restrict the port used at the application level but only at the OS level https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/troubleshooting.html#ip-ports
Suggestion to create a single NSG rule which will allows all TCP traffic on all ports within the compute subnet