Distributed (multi nodes) examples.

Example of distributed training

Prepare data.

Here we use ogbn_products and partition it into 2 partitions.

python partition_ogbn_dataset.py --dataset=ogbn-products --num_partitions=2

Launch training with:

2 nodes each with 2 GPUs

# node 0:
CUDA_VISIBLE_DEVICES=0,1 python dist_train_sage_supervised.py \
  --num_nodes=2 --node_rank=0 --master_addr=localhost

# node 1:
CUDA_VISIBLE_DEVICES=2,3 python dist_train_sage_supervised.py \
  --num_nodes=2 --node_rank=1 --master_addr=localhost

Note: you should change master_addr to the ip of node#0.

You can also use our script for the example:

Step 0: Setup a Distributed File System

Note: You may skip this step if you already set up folder(s) synchronized across machines.

To perform distributed sampling, files and codes need to be accessed across multiple machines. A distributed file system (i.e., NFS, Ceph, SSHFS, etc) exempt you from synchnonizing files such as partition information by hand.

For configuration details, please refer to the following links:

NFS: https://wiki.archlinux.org/index.php/NFS

SSHFS: https://www.digitalocean.com/community/tutorials/how-to-use-sshfs-to-mount-remote-file-systems-over-ssh

Ceph: https://docs.ceph.com/en/latest/install/

Step 1: Prepare data

Here we use ogbn_products and partition it into 2 partitions.

python partition_ogbn_dataset.py --dataset=ogbn-products --num_partitions=2

Step 2: Set up the configure file

An example template for configure file is in the yaml format. e.g. run_dist_train_sage_sup.yml.

# IP addresses for all nodes.
nodes:
  - 0.0.0.0
  - 1.1.1.1

# ssh IP ports for each node.
ports: [22, 22]

# path to python with GLT envs
python_bins:
  - /path/to/python
  - /path/to/python
# username for remote IPs.
usernames:
  - root
  - root

# dataset name, e.g. ogbn-products.
dataset: ogbn-products

# in_channel and out_channel for the dataset
in_channel: 100
out_channel: 47

node_ranks: [0, 1]

dst_paths:
  - /path/to/GLT
  - /path/to/GLT

# Setup visible cuda devices for each node.
visible_devices:
  - 0,1,2,3
  - 4,5,6,7

Step 3: Launch distributed jobs

pip install paramiko
pip install click
python run_dist_train_sage_sup.py --config=dist_train_sage_sup_config.yml --master_addr=0.0.0.0 --master_port=11234

Optional parameters which you can append after the command above includes:

--epochs: repeat epochs for sampling, (default=1)
--batch_size: batch size for sampling, (default=2048)

Here we present a comparison of end-to-end training time, as well as sampling and feature lookup time between GLT and DGL. We conducted the comparison using DGL version 0.9.1 and GLT version 0.2.0rc2 in an environment consisting of 2 nodes, each with 2 A100 GPUs.

Example of distributed training with server-client mode

The server-client mode enables the training to be decoupled, allowing for an arbitrary number of servers or clients. However, it's important to note that the number of clients must be equal to or greater than the number of servers, because each server needs at least one client to initiate it.

### Launch training with:
- 2 server nodes each with 1 server process for remote sampling and 2 GPUs.
- 3 client nodes each with 2 client processes for training and 2 GPUs.

server node 0:

CUDA_VISIBLE_DEVICES=0,1 python server_client_mode/sage_supervised_server.py
--num_server_nodes=2 --num_client_nodes=3 --node_rank=0 --num_dataset_partitions=2
--num_server_procs_per_node=1 --num_client_procs_per_node=2 --master_addr=localhost

server node 1:

CUDA_VISIBLE_DEVICES=2,3 python server_client_mode/sage_supervised_server.py
--num_server_nodes=2 --num_client_nodes=3 --node_rank=1 --num_dataset_partitions=2
--num_server_procs_per_node=1 --num_client_procs_per_node=2 --master_addr=localhost

client node 0:

CUDA_VISIBLE_DEVICES=4,5 python server_client_mode/sage_supervised_client.py
--num_server_nodes=2 --num_client_nodes=3 --node_rank=0 --num_dataset_partitions=3
--num_server_procs_per_node=1 --num_client_procs_per_node=2 --master_addr=localhost --pg_master=localhost

client node 1:

CUDA_VISIBLE_DEVICES=6,7 python server_client_mode/sage_supervised_client.py
--num_server_nodes=2 --num_client_nodes=3 --node_rank=1 --num_dataset_partitions=3
--num_server_procs_per_node=1 --num_client_procs_per_node=2 --master_addr=localhost --pg_master=localhost

client node 2:

CUDA_VISIBLE_DEVICES=8,9 python server_client_mode/sage_supervised_client.py
--num_server_nodes=2 --num_client_nodes=3 --node_rank=2 --num_dataset_partitions=3
--num_server_procs_per_node=1 --num_client_procs_per_node=2 --master_addr=localhost --pg_master=localhost

Note: 
1. You should change master_addr to the ip of server-node#0
2. The number of processes should be the same with the number of dataset partitions, i.e., num_server_procs_per_node * num_server_nodes = num_dataset_partitions.