We briefly discuss in the book the new class of thin client service characterized by OnLive and Gaikai. They require very little local client software, but at the expense that the images are computed remotely. Thus, the worst case end to end latency becomes the local latency (i.e. there is no longer any local control of camera direction, immediate actions such as firing etc.). This is the difference between locall end-to-end latency and distributed end-to-end latency (Paths A and D in Figure 10.1!).
They've used a simple video camera technique to measure latency, which we'll have more to say about in one of our first articles; the method is easy to reproduce but prone to some confounding factors. It is, however almost the best you can do for a game engine though it doesn't very accurately tell you about the network. A lot has been done in the VR community recently about more precise latency estimates, but they've used input systems that you have more control over.