Read/Write NetworkX Graph Object

Question

I am trying to deal with a super-massive NetworkX Graph object with hundreds of millions of nodes. I'd like to be able to write it to file as to not consume all my computer memory. However, I need to constantly be searching across existing nodes, updating edges, etc.

Is there a good solution for this? I'm not sure how it would work with any of the file formats provided on http://networkx.lanl.gov/reference/readwrite.html

The only solution i can think of is to store each node as a separate file with references to other nodes in the filesystem - that way, opening one node for examination doesn't overload the memory. Is there an existing filesystem for large amounts of data (e.g. PyTables) to do this without writing my own boilerplate code?

Answer 1

First try pickle; it's designed to serialize arbitrary objects.

An example of creating a DiGraph and serializing to a file:

import pickle
import networkx as nx

dg = nx.DiGraph()
dg.add_edge('a','b')
dg.add_edge('a','c')
pickle.dump(dg, open('/tmp/graph.txt', 'w'))

An example of loading a DiGraph from a file:

import pickle
import networkx as nx

dg = pickle.load(open('/tmp/graph.txt'))
print dg.edges()

Output:

[('a', 'c'), ('a', 'b')]

If this isn't efficient enough, I would write your own routine to serialize:

1. edges and
2. nodes (in case a node is incident to no edges).

Note that using list comprehensions when possible may be much more efficient (instead of standard for loops).

If this is not efficient enough, I'd call a C++ routine from within Python: http://docs.python.org/extending/extending.html

Answer 2

If you've built this as a NetworkX graph, then it will already be in memory. For this large of a graph, my guess is you'll have to do something similar to what you suggested with separate files. But, instead of using separate files, I'd use a database to store each node with many-to-many connections between nodes. In other words you'd have a table of nodes, and a table of edges, then to query for the neighbors of a particular node you could just query for any edges that have that particular node on either end. This should be fast, though I'm not sure if you'll be able to take advantage of NetworkX's analysis functions without first building the whole network in memory.