The network security people I know stopped bothering publishing attacks on I2P's directory service--which was(/is) in fact subject to eclipse attacks--because I2P didn't seem to care to fix the issues. Here's a great paper.
Great... are the issues getting fixed? The point I was making about I2P working so poorly that some people have given up trying to help--in addition to specifically noting that eclipse attacks do, in fact, affect the design of I2P--doesn't seem affected by "there exists people who still spend time demonstrating I2P doesn't work well" ;P.
I doubt the number of nodes is important when you can split a large group of them into a few subgroups and analyze them or alter how they interact with each other, then repeat subdivision; because again, them overlay networks are like smaller sandboxes inside the big one.
Yes, overlay network's configuration will be somewhat different, but physically it will be the same. It's like you're rearranging apple bits in a pie to save them; the pie will be sliced and eaten anyways.
