One aspect was the OLSR-NG project, which I started. OLSR.org was at that time programmed very inefficiently. We (Hannes, Henning, Bernd, me) re-wrote large parts of OLSR and now it uses a < 5% CPU load on 200MHz linksys devices in a network of 400 nodes.
Sorry, forgot to add: Sven-Ola, Markus and quite a few others.
The CPU improvements were dramatic! It makes sense to use trees instead of linked lists sometimes :) It makes sense, to use proper Dijkstra implementations instead of an O(n^2) implementation. So the main critique point against OLSR (high CPU usage) was actually proven to be wrong. It was simply the implementation which was broken (and got eventually fixed).
a.