2.2 Writing Efficient Message Passing Code


DGL optimizes memory consumption and computing speed for message passing. The optimization includes:

  • Merge multiple kernels in a single one: This is achieved by using update_all() to call multiple built-in functions at once. (Speed optimization)

  • Parallelism on nodes and edges: DGL abstracts edge-wise computation apply_edges() as a generalized sampled dense-dense matrix multiplication (gSDDMM) operation and parallelizes the computing across edges. Likewise, DGL abstracts node-wise computation update_all() as a generalized sparse-dense matrix multiplication (gSPMM) operation and parallelizes the computing across nodes. (Speed optimization)

  • Avoid unnecessary memory copy from nodes to edges: To generate a message that requires the feature from source and destination node, one option is to copy the source and destination node feature to that edge. For some graphs, the number of edges is much larger than the number of nodes. This copy can be costly. DGL’s built-in message functions avoid this memory copy by sampling out the node feature using entry index. (Memory and speed optimization)

  • Avoid materializing feature vectors on edges: the complete message passing process includes message generation, message aggregation and node update. In update_all() call, message function and reduce function are merged into one kernel if those functions are built-in. There is no message materialization on edges. (Memory optimization)

According to the above, a common practise to leverage those optimizations is to construct one’s own message passing functionality as a combination of update_all() calls with built-in functions as parameters.

For some cases like GATConv, where it is necessary to save message on the edges, one needs to call apply_edges() with built-in functions. Sometimes the messages on the edges can be high dimensional, which is memory consuming. DGL recommends keeping the dimension of edge features as low as possible.

Here’s an example on how to achieve this by splitting operations on the edges to nodes. The approach does the following: concatenate the src feature and dst feature, then apply a linear layer, i.e. \(W\times (u || v)\). The src and dst feature dimension is high, while the linear layer output dimension is low. A straight forward implementation would be like:

import torch
import torch.nn as nn

linear = nn.Parameter(torch.FloatTensor(size=(1, node_feat_dim * 2)))
def concat_message_function(edges):
     return {'cat_feat': torch.cat([edges.src.ndata['feat'], edges.dst.ndata['feat']])}
g.edata['out'] = g.edata['cat_feat'] * linear

The suggested implementation splits the linear operation into two, one applies on src feature, the other applies on dst feature. It then adds the output of the linear operations on the edges at the final stage, i.e. performing \(W_l\times u + W_r \times v\). This is because \(W \times (u||v) = W_l \times u + W_r \times v\), where \(W_l\) and \(W_r\) are the left and the right half of the matrix \(W\), respectively:

import dgl.function as fn

linear_src = nn.Parameter(torch.FloatTensor(size=(1, node_feat_dim)))
linear_dst = nn.Parameter(torch.FloatTensor(size=(1, node_feat_dim)))
out_src = g.ndata['feat'] * linear_src
out_dst = g.ndata['feat'] * linear_dst
g.srcdata.update({'out_src': out_src})
g.dstdata.update({'out_dst': out_dst})
g.apply_edges(fn.u_add_v('out_src', 'out_dst', 'out'))

The above two implementations are mathematically equivalent. The latter one is more efficient because it does not need to save feat_src and feat_dst on edges, which is not memory-efficient. Plus, addition could be optimized with DGL’s built-in function u_add_v, which further speeds up computation and saves memory footprint.