python - How do I run torch.distributed between Docker containers on separate instances using the bridge network? - Stack Overfl

admin管理员组

文章数量:1123706

I am trying to run a simple torch.distributed script between two Docker containers running on separate instances. Below is the code I am using:

import os
import torch
import torch.distributed as dist

def init_distributed():
    os.environ['MASTER_ADDR'] = "10.12.27.241"
    os.environ['MASTER_PORT'] = '29500'
    node_rank = int(os.environ.get('RANK', 0)) # 1 for worker
    world_size = 2

    dist.init_process_group(
        backend='gloo',
        rank=node_rank,
        world_size=world_size
    )
    print(f"Initialized process group: rank {node_rank} of {world_size}")
    return node_rank, world_size

def send_receive_message(rank, world_size):
    if rank == 0:
        # Node 0 sends a message
        message = torch.tensor([42, 43, 44], dtype=torch.int64)
        dist.send(message, dst=1)
        print(f"Rank {rank} sent message: {message}")
    else:
        # Node 1 receives the message
        message = torch.zeros(3, dtype=torch.int64)
        dist.recv(message, src=0)
        print(f"Rank {rank} received message: {message}")

if __name__ == "__main__":
    rank, world_size = init_distributed()
    send_receive_message(rank, world_size)
    # Barrier to ensure all processes have completed
    dist.barrier()
    # Clean up
    dist.destroy_process_group()

I am able to run this script successfully when using the --network=host option for docker run. However, due to organizational restrictions, I am required to use the --network=bridge option. When I use --network=bridge, I encounter the following error:

[E110 05:59:45.095859745 ProcessGroupGloo.cpp:143] Gloo connectFullMesh failed with [../third_party/gloo/gloo/transport/tcp/pair:144] no error
Traceback (most recent call last):
  File "/data/exp/com.py", line 36, in <module>
    rank, world_size = init_distributed()
  File "/data/exp/com.py", line 12, in init_distributed
    dist.init_process_group(
  File "/usr/local/lib/python3.10/dist-packages/torch/distributed/c10d_logger.py", line 83, in wrapper
    return func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/distributed/c10d_logger.py", line 97, in wrapper
    func_return = func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/distributed/distributed_c10d.py", line 1527, in init_process_group
    default_pg, _ = _new_process_group_helper(
  File "/usr/local/lib/python3.10/dist-packages/torch/distributed/distributed_c10d.py", line 1744, in _new_process_group_helper
    backend_class = ProcessGroupGloo(
RuntimeError: Gloo connectFullMesh failed with [../third_party/gloo/gloo/transport/tcp/pair:144] no error

How can I configure torch.distributed to work with the bridge network when running containers on separate instances? What additional steps or configurations are required to make Gloo backend communication succeed in this setup?

Any guidance or pointers would be greatly appreciated!

本文标签：