Menu
I was to set up DDP (distributed data parallel) on a DGX A100 but it doesn't work. Whenever I try to run it simply hangs. My code is super simple just spawning 4 processes for 4 gpus (for the sake of debugging I simply destroy the group immediately but it doesn't even reach there):
def find_free_port():
""" https://stackoverflow.com/questions/1365265/on-localhost-how-do-i-pick-a-free-port-number """
import socket
from contextlib import closing
with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as s:
s.bind(('', 0))
s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
return str(s.getsockname()[1])
def setup_process(rank, world_size, backend='gloo'):
"""
Initialize the distributed environment (for each process).
gloo: is a collective communications library (https://github.com/facebookincubator/gloo). My understanding is that
it's a library/API for process to communicate/coordinate with each other/master. It's a backend library.
export NCCL_SOCKET_IFNAME=eth0
export NCCL_IB_DISABLE=1
https://stackoverflow.com/questions/61075390/about-pytorch-nccl-error-unhandled-system-error-nccl-version-2-4-8
https://pytorch.org/docs/stable/distributed.html#common-environment-variables
"""
if rank != -1: # -1 rank indicates serial code
print(f'setting up rank={rank} (with world_size={world_size})')
# MASTER_ADDR = 'localhost'
MASTER_ADDR = '127.0.0.1'
MASTER_PORT = find_free_port()
# set up the master's ip address so this child process can coordinate
os.environ['MASTER_ADDR'] = MASTER_ADDR
print(f"{MASTER_ADDR=}")
os.environ['MASTER_PORT'] = MASTER_PORT
print(f"{MASTER_PORT}")
# - use NCCL if you are using gpus: https://pytorch.org/tutorials/intermediate/dist_tuto.html#communication-backends
if torch.cuda.is_available():
# unsure if this is really needed
# os.environ['NCCL_SOCKET_IFNAME'] = 'eth0'
# os.environ['NCCL_IB_DISABLE'] = '1'
backend = 'nccl'
print(f'{backend=}')
# Initializes the default distributed process group, and this will also initialize the distributed package.
dist.init_process_group(backend, rank=rank, world_size=world_size)
# dist.init_process_group(backend, rank=rank, world_size=world_size)
# dist.init_process_group(backend='nccl', init_method='env://', world_size=world_size, rank=rank)
print(f'--> done setting up rank={rank}')
dist.destroy_process_group()
mp.spawn(setup_process, args=(4,), world_size=4)
why is this hanging?
nvidia-smi output:
$ nvidia-smi
Fri Mar 5 12:47:17 2021
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.102.04 Driver Version: 450.102.04 CUDA Version: 11.0 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 A100-SXM4-40GB On | 00000000:07:00.0 Off | 0 |
| N/A 26C P0 51W / 400W | 0MiB / 40537MiB | 0% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
| 1 A100-SXM4-40GB On | 00000000:0F:00.0 Off | 0 |
| N/A 25C P0 52W / 400W | 3MiB / 40537MiB | 0% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
| 2 A100-SXM4-40GB On | 00000000:47:00.0 Off | 0 |
| N/A 25C P0 51W / 400W | 3MiB / 40537MiB | 0% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
| 3 A100-SXM4-40GB On | 00000000:4E:00.0 Off | 0 |
| N/A 25C P0 51W / 400W | 3MiB / 40537MiB | 0% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
| 4 A100-SXM4-40GB On | 00000000:87:00.0 Off | 0 |
| N/A 30C P0 52W / 400W | 3MiB / 40537MiB | 0% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
| 5 A100-SXM4-40GB On | 00000000:90:00.0 Off | 0 |
| N/A 29C P0 53W / 400W | 0MiB / 40537MiB | 0% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
| 6 A100-SXM4-40GB On | 00000000:B7:00.0 Off | 0 |
| N/A 29C P0 52W / 400W | 0MiB / 40537MiB | 0% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
| 7 A100-SXM4-40GB On | 00000000:BD:00.0 Off | 0 |
| N/A 48C P0 231W / 400W | 7500MiB / 40537MiB | 99% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 7 N/A N/A 147243 C python 7497MiB |
+-----------------------------------------------------------------------------+
How do I set up ddp in this new machine?
btw I've successfully installed APEX because some other links say to do that but it still fails. For I did:
went to: https://github.com/NVIDIA/apex follwed their instructions
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
but before the above I had to update gcc:
conda install -c psi4 gcc-5
it did install it as I successfully imported it but it didn't help.
Now it actually prints an error msg:
Traceback (most recent call last):
File "/home/miranda9/miniconda3/envs/metalearning/lib/python3.8/multiprocessing/process.py", line 315, in _bootstrap
self.run()
File "/home/miranda9/miniconda3/envs/metalearning/lib/python3.8/multiprocessing/process.py", line 108, in run
self._target(*self._args, **self._kwargs)
File "/home/miranda9/miniconda3/envs/metalearning/lib/python3.8/site-packages/torch/multiprocessing/spawn.py", line 19, in _wrap
fn(i, *args)
KeyboardInterrupt
Process SpawnProcess-3:
Traceback (most recent call last):
File "/home/miranda9/miniconda3/envs/metalearning/lib/python3.8/site-packages/torch/multiprocessing/spawn.py", line 19, in _wrap
fn(i, *args)
File "/home/miranda9/ML4Coq/ml4coq-proj/embeddings_zoo/tree_nns/main_brando.py", line 252, in train
setup_process(rank, world_size=opts.world_size)
File "/home/miranda9/ML4Coq/ml4coq-proj/embeddings_zoo/distributed.py", line 85, in setup_process
dist.init_process_group(backend, rank=rank, world_size=world_size)
File "/home/miranda9/miniconda3/envs/metalearning/lib/python3.8/site-packages/torch/distributed/distributed_c10d.py", line 436, in init_process_group
store, rank, world_size = next(rendezvous_iterator)
File "/home/miranda9/miniconda3/envs/metalearning/lib/python3.8/site-packages/torch/distributed/rendezvous.py", line 179, in _env_rendezvous_handler
store = TCPStore(master_addr, master_port, world_size, start_daemon, timeout)
RuntimeError: connect() timed out.
During handling of the above exception, another exception occurred:
related:
734
The following fixes are based on Writing Distributed Applications with PyTorch, Initialization Methods.
Issue 1:
It will hang unless you pass in nprocs=world_size to mp.spawn(). In other words, it's waiting for the "whole world" to show up, process-wise.
Issue 2:
The MASTER_ADDR and MASTER_PORT need to be the same in each process' environment and need to be a free address:port combination on the machine where the process with rank 0 will be run.
Both of these are implied or directly read from the following quote from the link above (emphasis added):
Environment Variable
We have been using the environment variable initialization method throughout this tutorial. By setting the following four environment variables on all machines, all processes will be able to properly connect to the master, obtain information about the other processes, and finally handshake with them.
MASTER_PORT: A free port on the machine that will host the process with rank 0.
MASTER_ADDR: IP address of the machine that will host the process with rank 0.
WORLD_SIZE: The total number of processes, so that the master knows how many workers to wait for.
RANK: Rank of each process, so they will know whether it is the master of a worker.
Here's some code to demonstrate both of those in action:
import torch
import torch.multiprocessing as mp
import torch.distributed as dist
import os
def find_free_port():
""" https://stackoverflow.com/questions/1365265/on-localhost-how-do-i-pick-a-free-port-number """
import socket
from contextlib import closing
with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as s:
s.bind(('', 0))
s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
return str(s.getsockname()[1])
def setup_process(rank, master_addr, master_port, world_size, backend='gloo'):
print(f'setting up {rank=} {world_size=} {backend=}')
# set up the master's ip address so this child process can coordinate
os.environ['MASTER_ADDR'] = master_addr
os.environ['MASTER_PORT'] = master_port
print(f"{master_addr=} {master_port=}")
# Initializes the default distributed process group, and this will also initialize the distributed package.
dist.init_process_group(backend, rank=rank, world_size=world_size)
print(f"{rank=} init complete")
dist.destroy_process_group()
print(f"{rank=} destroy complete")
if __name__ == '__main__':
world_size = 4
master_addr = '127.0.0.1'
master_port = find_free_port()
mp.spawn(setup_process, args=(master_addr,master_port,world_size,), nprocs=world_size)
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With
