TCP port collision when running two different tests on the same machine

[jamesETsmith]

🔨Work Item

When running two separate test runs on the same machine, the tests can crash because they are both trying to use the same tcp port. See

 rg "tcp://127.0.0.1:12345"
...
tests/python/pytorch/graphbolt/test_item_sampler.py
886:        else "tcp://127.0.0.1:12345"

tests/python/pytorch/graphbolt/test_dataloader.py
92:            else "tcp://127.0.0.1:12345"

tests/python/pytorch/cuda/test_nccl.py
17:        init_method="tcp://127.0.0.1:12345",
40:        init_method="tcp://127.0.0.1:12345",
63:        init_method="tcp://127.0.0.1:12345",
89:        init_method="tcp://127.0.0.1:12345",
...

Description

We should have the tcp port be some random number within a certain range to ensure that TCP port collision is really unlikely.

[GMNGeoffrey]

You might want to see if you can bind to an ephemeral port and have the os allocate an unused one for you. This is usually done by using port 0. Not sure if tcp://127.0.0.1:0 would achieve this.

[jamesETsmith]

@GMNGeoffrey that's a good idea, I'll start there

[github-actions]

This issue has been automatically marked as stale due to lack of activity. It will be closed if no further activity occurs. Thank you