In my experience, most time in Tramp is spent on wire. That is, sending
commands to the remote shell, and reading the command's output. Compared
with this, time spent in Elisp is not so important (although any
optimization will be appreciated as well).
Based on this, I wonder how useful/difficult it would be to add a sort of "ping timer" to buffers that make regular use of remote commands, measuring the delta between just before the most recent send and the following (assuming no interleaving) recv. With that raw data, it should be relatively easy to add things like a moving average and/or an alert for sudden spikes.
Would this be more likely to result in a useful diagnostic or a computed indicator that slow things are slow?
~Chad