I'm working with a graph that has about 16 million vertices and 320 million edges.  Currently, I'm using graph-tool in Python but building the graph is far too slow.  Before I start porting to C++ I starting reading up on the Parallel Boost Graph Library.  From the examples in the documentation, it seems that one must build the graph on one node in the cluster, then call synchronize to send that graph to all nodes.  Is there a way to build parts of the graph on each node and then synchronize those parts?

Some background: the graph actually comes from a CT scan.  The vertices are voxels, the edges are relationships between adjacent voxels.  At the moment I'm working on a very small test volume - 256^3 voxels - but later volumes will have billions of vertices and edges.

Thanks for the help.