Distributed Node Classification

In this tutorial, we will walk through the steps of performing distributed GNN training for a node classification task. To understand distributed GNN training, you need to read the tutorial of multi-GPU training first. This tutorial is developed on top of multi-GPU training by providing extra steps for partitioning a graph, modifying the training script and setting up the environment for distributed training.

Partition a graph

In this tutorial, we will use OGBN products graph as an example to illustrate the graph partitioning. Let’s first load the graph into a DGL graph. Here we store the node labels as node data in the DGL Graph.

import os
os.environ['DGLBACKEND'] = 'pytorch'
import dgl
import torch as th
from ogb.nodeproppred import DglNodePropPredDataset
data = DglNodePropPredDataset(name='ogbn-products')
graph, labels = data[0]
labels = labels[:, 0]
graph.ndata['labels'] = labels

We need to split the data into training/validation/test set during the graph partitioning. Because this is a node classification task, the training/validation/test sets contain node IDs. We recommend users to convert them as boolean arrays, in which True indicates the existence of the node ID in the set. In this way, we can store them as node data. After the partitioning, the boolean arrays will be stored with the graph partitions.

splitted_idx = data.get_idx_split()
train_nid, val_nid, test_nid = splitted_idx['train'], splitted_idx['valid'], splitted_idx['test']
train_mask = th.zeros((graph.num_nodes(),), dtype=th.bool)
train_mask[train_nid] = True
val_mask = th.zeros((graph.num_nodes(),), dtype=th.bool)
val_mask[val_nid] = True
test_mask = th.zeros((graph.num_nodes(),), dtype=th.bool)
test_mask[test_nid] = True
graph.ndata['train_mask'] = train_mask
graph.ndata['val_mask'] = val_mask
graph.ndata['test_mask'] = test_mask

Then we call the partition_graph function to partition the graph with METIS and save the partitioned results in the specified folder. Note: partition_graph runs on a single machine with a single thread. You can go to our user guide to see more information on distributed graph partitioning.

The code below shows an example of invoking the partitioning algorithm and generate four partitions. The partitioned results are stored in a folder called 4part_data. While partitioning a graph, we allow users to specify how to balance the partitions. By default, the algorithm balances the number of nodes in each partition as much as possible. However, this balancing strategy is not sufficient for distributed GNN training because some partitions may have many more training nodes than other partitions or some partitions may have more edges than others. As such, partition_graph provides two additional arguments balance_ntypes and balance_edges to enforce more balancing criteria. For example, we can use the training mask to balance the number of training nodes in each partition, as shown in the example below. We can also turn on the balance_edges flag to ensure that all partitions have roughly the same number of edges.

dgl.distributed.partition_graph(graph, graph_name='ogbn-products', num_parts=4,
                                out_path='4part_data',
                                balance_ntypes=graph.ndata['train_mask'],
                                balance_edges=True)

When partitioning a graph, DGL shuffles node IDs and edge IDs so that nodes/edges assigned to a partition have contiguous IDs. This is necessary for DGL to maintain the mappings of global node/edge IDs and partition IDs. If a user needs to map the shuffled node/edge IDs to their original IDs, they can turn on the return_mapping flag of partition_graph, which returns a vector for the node ID mapping and edge ID mapping. Below shows an example of using the ID mapping to save the node embeddings after distributed training. This is a common use case when users want to use the trained node embeddings in their downstream task. Below let’s assume that the trained node embeddings are stored in the node_emb tensor, which is indexed by the shuffled node IDs. We shuffle the embeddings again and store them in the orig_node_emb tensor, which is indexed by the original node IDs.

nmap, emap = dgl.distributed.partition_graph(graph, graph_name='ogbn-products',
                                             num_parts=4,
                                             out_path='4part_data',
                                             balance_ntypes=graph.ndata['train_mask'],
                                             balance_edges=True,
                                             return_mapping=True)
orig_node_emb = th.zeros(node_emb.shape, dtype=node_emb.dtype)
orig_node_emb[nmap] = node_emb

Distributed training script

The distributed training script is very similar to multi-GPU training script with just a few modifications. It also relies on the Pytorch distributed component to exchange gradients and update model parameters. The distributed training script only contains the code of the trainers.

Initialize network communication

Distributed GNN training requires to access the partitioned graph structure and node/edge features as well as aggregating the gradients of model parameters from multiple trainers. DGL’s distributed component is responsible for accessing the distributed graph structure and distributed node features and edge features while Pytorch distributed is responsible for exchanging the gradients of model parameters. As such, we need to initialize both DGL and Pytorch distributed components at the beginning of the training script.

We need to call DGL’s initialize function to initialize the trainers’ network communication and connect with DGL’s servers at the very beginning of the distributed training script. This function has an argument that accepts the path to the cluster configuration file.

import dgl
import torch as th
dgl.distributed.initialize(ip_config='ip_config.txt')

The configuration file ip_config.txt has the following format:

ip_addr1 [port1]
ip_addr2 [port2]

Each row is a machine. The first column is the IP address and the second column is the port for connecting to the DGL server on the machine. The port is optional and the default port is 30050.

After initializing DGL’s network communication, a user can initialize Pytorch’s distributed communication.

th.distributed.init_process_group(backend='gloo')

Reference to the distributed graph

DGL’s servers load the graph partitions automatically. After the servers load the partitions, trainers connect to the servers and can start to reference to the distributed graph in the cluster as below.

g = dgl.distributed.DistGraph('ogbn-products')

As shown in the code, we refer to a distributed graph by its name. This name is basically the one passed to the partition_graph function as shown in the section above.

Get training and validation node IDs

For distributed training, each trainer can run its own set of training nodes. The training nodes of the entire graph are stored in a distributed tensor as the train_mask node data, which was constructed before we partitioned the graph. Each trainer can invoke node_split to its set of training nodes. The node_split function splits the full training set evenly and returns the training nodes, majority of which are stored in the local partition, to ensure good data locality.

train_nid = dgl.distributed.node_split(g.ndata['train_mask'])

We can split the validation nodes in the same way as above. In this case, each trainer gets a different set of validation nodes.

valid_nid = dgl.distributed.node_split(g.ndata['val_mask'])

Define a GNN model

For distributed training, we define a GNN model exactly in the same way as mini-batch training or full-graph training. The code below defines the GraphSage model.

import torch.nn as nn
import torch.nn.functional as F
import dgl.nn as dglnn
import torch.optim as optim

class SAGE(nn.Module):
    def __init__(self, in_feats, n_hidden, n_classes, n_layers):
        super().__init__()
        self.n_layers = n_layers
        self.n_hidden = n_hidden
        self.n_classes = n_classes
        self.layers = nn.ModuleList()
        self.layers.append(dglnn.SAGEConv(in_feats, n_hidden, 'mean'))
        for i in range(1, n_layers - 1):
            self.layers.append(dglnn.SAGEConv(n_hidden, n_hidden, 'mean'))
        self.layers.append(dglnn.SAGEConv(n_hidden, n_classes, 'mean'))

    def forward(self, blocks, x):
        for l, (layer, block) in enumerate(zip(self.layers, blocks)):
            x = layer(block, x)
            if l != self.n_layers - 1:
                x = F.relu(x)
        return x

num_hidden = 256
num_labels = len(th.unique(g.ndata['labels'][0:g.num_nodes()]))
num_layers = 2
lr = 0.001
model = SAGE(g.ndata['feat'].shape[1], num_hidden, num_labels, num_layers)
loss_fcn = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=lr)

For distributed training, we need to convert the model into a distributed model with Pytorch’s DistributedDataParallel.

model = th.nn.parallel.DistributedDataParallel(model)

Distributed mini-batch sampler

We can use the same DistNodeDataLoader, the distributed counterpart of NodeDataLoader, to create a distributed mini-batch sampler for node classification.

sampler = dgl.dataloading.MultiLayerNeighborSampler([25,10])
train_dataloader = dgl.dataloading.DistNodeDataLoader(
                             g, train_nid, sampler, batch_size=1024,
                             shuffle=True, drop_last=False)
valid_dataloader = dgl.dataloading.DistNodeDataLoader(
                             g, valid_nid, sampler, batch_size=1024,
                             shuffle=False, drop_last=False)

Training loop

The training loop for distributed training is also exactly the same as the single-process training.

import sklearn.metrics
import numpy as np

for epoch in range(10):
    # Loop over the dataloader to sample mini-batches.
    losses = []
    with model.join():
        for step, (input_nodes, seeds, blocks) in enumerate(train_dataloader):
            # Load the input features as well as output labels
            batch_inputs = g.ndata['feat'][input_nodes]
            batch_labels = g.ndata['labels'][seeds]

            # Compute loss and prediction
            batch_pred = model(blocks, batch_inputs)
            loss = loss_fcn(batch_pred, batch_labels)
            optimizer.zero_grad()
            loss.backward()
            losses.append(loss.detach().cpu().numpy())
            optimizer.step()

    # validation
    predictions = []
    labels = []
    with th.no_grad(), model.join():
        for step, (input_nodes, seeds, blocks) in enumerate(valid_dataloader):
            inputs = g.ndata['feat'][input_nodes]
            labels.append(g.ndata['labels'][seeds].numpy())
            predictions.append(model(blocks, inputs).argmax(1).numpy())
        predictions = np.concatenate(predictions)
        labels = np.concatenate(labels)
        accuracy = sklearn.metrics.accuracy_score(labels, predictions)
        print('Epoch {}: Validation Accuracy {}'.format(epoch, accuracy))

Set up distributed training environment

After partitioning a graph and preparing the training script, we now need to set up the distributed training environment and launch the training job. Basically, we need to create a cluster of machines and upload both the training script and the partitioned data to each machine in the cluster. A recommended solution of sharing the training script and the partitioned data in the cluster is to use NFS (Network File System).

For any users who are not familiar with NFS, below is a small tutorial of setting up NFS in an existing cluster.

NFS server side setup (ubuntu only)

First, install essential libs on the storage server

sudo apt-get install nfs-kernel-server

Below we assume the user account is ubuntu and we create a directory of workspace in the home directory.

mkdir -p /home/ubuntu/workspace

We assume that the all servers are under a subnet with ip range 192.168.0.0 to 192.168.255.255. We need to add the following line to /etc/exports

/home/ubuntu/workspace  192.168.0.0/16(rw,sync,no_subtree_check)

Then restart NFS, the setup on server side is finished.

sudo systemctl restart nfs-kernel-server

For configuration details, please refer to NFS ArchWiki (https://wiki.archlinux.org/index.php/NFS).

NFS client side setup (ubuntu only)

To use NFS, clients also require to install essential packages

sudo apt-get install nfs-common

You can either mount the NFS manually

mkdir -p /home/ubuntu/workspace
sudo mount -t nfs <nfs-server-ip>:/home/ubuntu/workspace /home/ubuntu/workspace

or add the following line to /etc/fstab so the folder will be mounted automatically

<nfs-server-ip>:/home/ubuntu/workspace   /home/ubuntu/workspace   nfs   defaults    0 0

Then run

mount -a

Now go to /home/ubuntu/workspace and save the training script and the partitioned data in the folder.

SSH Access

The launch script accesses the machines in the cluster via SSH. Users should follow the instruction in this document to set up the passwordless SSH login on every machine in the cluster. After setting up the passwordless SSH, users need to authenticate the connection to each machine and add their key fingerprints to ~/.ssh/known_hosts. This can be done automatically when we ssh to a machine for the first time.

Launch the distributed training job

After everything is ready, we can now use the launch script provided by DGL to launch the distributed training job in the cluster. We can run the launch script on any machine in the cluster.

python3 ~/workspace/dgl/tools/launch.py   --workspace ~/workspace/   --num_trainers 1   --num_samplers 0   --num_servers 1   --part_config 4part_data/ogbn-products.json   --ip_config ip_config.txt   "python3 train_dist.py"

If we split the graph into four partitions as demonstrated at the beginning of the tutorial, the cluster has to have four machines. The command above launches one trainer and one server on each machine in the cluster. ip_config.txt lists the IP addresses of all machines in the cluster as follows:

ip_addr1
ip_addr2
ip_addr3
ip_addr4