Let's start to think laterally about this. Many of the problems around CCID3/TFRC implementation seem to be on local LANs and rtt is less than t_gran.
I would agree that this is a problem. One thing that should help would be to modify Section 4.7 of rfc3448bis, on "Scheduling of Packet Transmissions", to say that instead of allowing packets to be sent t_delta seconds before their nominal send time, for: t_delta = min(t_ipi, t_gran)/2 we should instead recommend that t_delta is also at most rtt/2 (or some other fraction of the rtt); t_delta = min(t_ipi, t_gran, rtt)/2. Then when rtt is less than t_gran, the sender would not be sending more that an RTT of packets before their nominal send time. When this is combined with the guideline that the sender can maintain unused send credits up to an RTT, I think that this gives behavior roughly comparable to that of TCP. (I agree, however, that the performance of both TCP and TFRC in environments with very small round-trip times could use further investigation.) - Sally http://www.icir.org/floyd/
