[e2e] congestion control for flows w. small packets

[Joerg Widmer]

Hi,
I'd like to hear comments on what you think is an acceptable congestion
control behavior for flows that send small packets (streaming audio
etc.).

TCP throughput scales linearly with the packet size (half the packet
size results in half the throughput). This behavior is reasonable if 
packets "cost" the same indepently of their size. However, if the
only scarce resource in the network is bandwidth, then it makes sense to
send with the same bitrate as a TCP flow with MTU sized packets but with
a higher packet rate in case the own packets are smaller. I guess a flow
that "puts the same overall burden" on the network as a TCP flow with
normal-sized packets will have a rate somewhere inbetween those
two choices (e.g., something like no more throughput than TCP with
min(0.5 MTU,576) byte packets under the same networks conditions).

To be able to achieve a rate equivalent to TCP with a certain packet
size, we need to know how sending smaller packets influences the loss
rate a flow experiences. With drop-tail queues or with RED in packet
mode (see Sally's http://www.aciri.org/floyd/REDaveraging.txt) the
packet drop probability is relatively independent of the packet size. To
estimate the loss rate a flow with larger packets would experience we
can simply aggregate arrived packets (count 10 arrived packets as one)
as well as lost packets (count 10 lost packets as one lost packet).
However, if the drop probability depends on the packet size such as in
RED in byte mode, we should only aggregate arrived packets but not
packet loss.

If queue management and aggregation scheme don't match, we have two
cases:
- if queuing is RED in byte mode but we aggregate loss the flow with
small packets will estimate a far too low loss rate
- if queuing is drop-tail or RED in packet mode and we don't aggregate
loss the flow will estimate a far too high loss rate (and everybody
would just continue to use UDP w/o congestion control since the
resulting throughput would be too low to be useful)

If they do match, aggregation works very well.

Is it acceptable to be too aggressive in case the packet drop
probability depends on the packet size? Do any networks have packet drop
probabilities dependent on the packet size?

Thanks,
  Joerg

P.S. I ran a number of simulations on this and I'm happy to send the
results to whoever is interested.