We're trying to measure runtime overheads, not mapper overheads. Furthermore, we need to make sure we're getting the best possible mapping for each application and machine, and the only way I'm going to believe any Legion numbers is if someone has hand written an implementation for every mapper function for each application on every machine. These are the kinds of applications where getting absolutely a perfect mapping is crucial for performance and the default mapper's heuristics are not going to get them right. The whole point of having a mapper interface is so we don't have to rely on the heuristics of a runtime implementation. Therefore all Legion applications need multiple custom mappers that compute perfect mappings on every machine.