1) Eric talked about his idea of measuring the dynamics of short lived
TCP connections under varying inelastic contention. The idea is to send
enough inelastic traffic (e.g. UDP streams) to fill up most of the
queues at some nearby bottleneck router and see how measured short-lived
TCP flows adapt (esp with all the other TCP's there as well). I
believe this is an approximation to underbuffering.
2) We also talked about John's idea of sending more packets (e.g. 3) on
startup to avoid the cost of slow start. We agreed this didn't matter
much for modem links but might for show transaction of high bW*delay
links. John gave the argument that under conditions of random queue
occupancy that it might be better to put all of the traffic together so
you get an all or nothing loss event. The opposing logic is that
increasing the opening window size increases the loss probability. It
was suggested that we could measure the amount of traffic that grows to
ssthresh=3 packets without dropping to get a sense for whether there is
enough buffer space to allow this.
3) ACK aggregation was suggested for the detour network. We could put
many ACK's together in one packet to decrease queue occupancy. I didn't
say it during the meeting but there is an obvious downside to doing this
which is that the loss of that single packet has serious effects for
everyone. It might make sense with the ACK priority scheme I suggested.
Send out your suggests for interesting show projects/studies and I'll
compile them.
- Stefan