Really, consider a link where the path carrying most of the data has a time of N and the path carrying the ack has a path 2N or N + 1/2N You are computing an average which seems like it could be skewed based on the path the ack takes.
This doesn't matter. The key notions for TCP are (1) what's the time scale for feedback, and (2) what's the pipe size (= bandwidth-delay product). Both of these are determined by the RTT and not the one-way prop time.
Does anyone know of any papers on the effect of asymetric paths on TCP performance?
Well, I thought about this quite a bit for my end-to-end routing study (ftp://ftp.ee.lbl.gov/papers/routing.SIGCOMM.ps.Z). I really wanted to come up with some reason why asymmetric routing has serious implications for TCP performance, but wasn't able to. I guess this is a good thing, since 50% of the paths in my study were asymmetric in terms of visiting at least one different city in the two directions. 30% visited at least one different AS. Vern