[Brad Abrams] Now normally in the case I ask a simple question: Have you measured it? Measurement is at the heart of perf work. Random speculation is not (which is why I could never be a good perf guy;-)). In many cases the answer is no and in fact the api is only ever called as part of a blocking network IO operation so the extra couple of milliseconds is pretty much noise. Or something like that. And in general I don’t think that loading one additional type is a big thing to worry about even if it was happening…

The basic rules to follow: Explore, understand what API’s are doing under the hood, use tools (CLR Profiler, Reflector) and measure everything.