raj@hpindwa.cup.hp.com (Rick Jones) (03/01/91)
A few toss-up questions: Does anyone have data showing the 'hit rate' of pulling as much data as possible into a retransmission? Put another way, how often does a series of packets get lost vs a single packet? I think the effectiveness is rather low, but am not sure. This comes out of some 'deep-thought' I've been trying to do concerning BSD's implementing the congestion control packet windows as byte windows - presumeably because BSD does not have a concept of a 'packet rtxq' in favor of a 'bytes rtxq' (Does that change in 4.4?) In another life ;-) I worked-on a system where there was a 'packet rtxq', which would seem to more purely implement a packet congestion window scheme (primarily because it knows how many 'packets' are out there and doesn't estimate it using MSS's). The trade-off (which I think was unconscious) was that there was no 'pulling-up' of data into retransissions. If a 10 byte packet was lost, then a 10 byte packet would be retransmitted. It also seems to have more frequently achieved exponential growth in packet window, whereas BSD seems to effectively get linear (one MSS per ACK packet - or does that change too?) Hence my question as to the effectiveness of pulling-up data into rtxs. There are other questions, but I'll float these first. rick jones