Resource discovery is an important aspect of many modern large-scale distributed systems. In the past, this problem has been solved using many different approaches, such as a central registry server, flooding-based protocols, and distributed hash tables. In this paper, these three widely used architectures are compared, using measurement results obtained from real implementations run on an Emulab emulation environment. This allows us to study the advantages and disadvantages of the architectures and determine their usefulness. The measurement study lead to several interesting conclusions. First, the centralised architecture incurs the least traffic overhead. However, it balances the load poorly, and introduces a single point-of-failure. Second, of the two decentralised architectures, the distributed hash table generates the least overhead. Finally, hierarchical architectures were shown to be most effective when the fraction of super-peers compared to regular peers is small.