Perhaps the distribution is multimodal, i.e. the distributions could be a sum of different conditions and hence multiple distributions, e.g. such as a system that is idle vs. a system that is doing heavy single file downloading vs. a system that is streaming requiring regular intervals of high data throughput. The time spaces between pings may also be important, and some Nyquist-limited information may be lost in inadequate sampling.
There is an additional post here about summing multiple exponent distributions. I could not comment on nor upvote that one, so I made this separate response. The point on multiple hops is very valid, and I would first try analyzing pings to the averagingnext immediate node.