I have a graph that I want to execute a breadth-first search (BFS) on.  When each node is visited, it spawns a thread to execute some work.  I must not execute a child node before the parent execution is completed, but I also want to maximize concurrency. In other words, I don't want to "serialize" the BFS to force proper execution ordering.

Are there any ideas of how to accomplish this efficiently?

James