its would certainly be straightforward to show the standard simulation
small-number-of-steady-state-tcps dynamics for sample networks in each
of these regimes. but I presume you're looking to come up with a
stronger result. we could also do a comparitive study of the
requirements of each of the architectures. that is a prerequisite for
such a paper, but isn't sufficient to make it interesting. I'll try to
think about what we could hope to show in the area, but hopefully
you'll write back if you already have a plan.
But I believe we could demonstrate that no fixed parameter
choices work well in all environments -- eg, what would work
well in a WAN for large RTTs and high drops won't work as well in LANs.
anecdotal evidence (John Hawkinson of BBN) suggests that in fact the
overall behaviour of RED (I don't know exactly what behaviours we was
describing) is fairly sensitive to the capacity of the links, queue
depths, and the RTT of the flows
And as a final thought, one of the papers in my queue suggested
that you don't need per-flow credits, that you can get by with
per-destination credits, since you can multiplex together the
packets for the same destination.
there is actually a fair amount of flexibility concerning how you
decide to aggregate a credit mechanism. clearly higher aggregates mean
less credit exchange, a smaller number of buffers across all the
aggregates at intermediate nodes, and less state at the intermediate
nodes.
the real problem is deciding what the correct dynamic credit
allocation policy is, especially given the dynamic nature (short
flows, or bursty numbers of flows in the aggregate) of the traffic
I still think (without evidence) that rate-based credits are worth
looking into just by the nature of having less control overhead and
being more composable along a path. were there other reasons aside
from an absolute no-drop policy that the digital switches moved from
rate to buffer based credits?