I've deceided to use the delta times between TCP SYN Ethernet frames to determin network latency. I've collected this data over time. This now presents a philosopical question; i.e., what exactly is network latency? One could argue that a network's baseline latency is the fastest time seen. Others might think that the average of these discovered delta times would be a network's latency. I'm wondering what folks think about that.
There are a couple of ways to look at a system. Some would simply look at delay as a function of frequency and that would tell the whole story. If the signals are bunched into a pass band, then the delay at that frequency, the group delay, would be a central measure. Note that this says little more about the system. So, if the communication channel were also band-limited then there could also be something to be said for the rate of change of the signals in that band. Now, in systems where we have a guaranteed signalling bandwidth - like Ethernet perhaps - we might also talk about bandwidth and latency. This is because (as also above) the bandwidth measure alone doesn't tell us anything about delay / latency. So, as you know, the latency becomes an important factor if you are dealing perhaps with real time systems, etc.....
So far I've lumped delay and latency together - as in linear systems theory. However, latency generally contains more than simply the delay of a linear system. It includes the delays that are caused by processing. As such, latency may be a variable as you've considered. In that case I believe the engineering measure for latency would come in two flavors depending on your application reuirements:
- the worst-case (largest or smallest) latency is important in things like interrupt-driven real time systems. So, you have to pick the largest number that you see in order to guarantee performance in a design or to predict what can happen in doing an analysis. The shortest latency may also be used in a worst-case analysis if your purpose is to rely on latency for some system characteristic. I should think this would be an unusual case but mention it for completeness.
- the average latency may be of some value if your analysis is statistical.
So, you have to answer your own question it appears. I would guess that most often, and for most practical purposes, the network latency is the *largest delay* that can be measured or the largest delay with 95% (pick your number) probability of occurrence over a given time, etc. The latter might be useful if you want to know how bad things can be "most of the time", allowing for an occasional event to be even worse.