Re: TCP and Loss (inherently lossy nets)


WESTCOTT@G.BBN.COM
7 Oct 1987 14:58-EDT


Mike,

Hop by hop retransmissions are necessary in packet radio networks
because PRs must transmit over links with relatively low probability
of success. Imagine a 5 hop path with a probability of 80% success
per hop (~30% over net). When that is coupled with congestion and other
problems along a longer Internet route, then it is extremely difficult
and costly (in terms of packet transmissions) to attempt to maintain
a reliable connection.

Presently, PR nets also use the timing information from their hop
acknowledgements to apply backpressure for congestion control; its
known as the pacing algorithm. The reasoning is that rather than fill
all the buffers in a data path, limit the packets forwarded to what
the next PR can handle. If congestion appears on the far side of the
net, traffic generation may be slowed by increasing the pacing delays
all the way back to the source. Smart sources slow down, dumb sources
get their packets dropped by the source's attached PR.

Hop by hop retransmission is limited to several attempts along the
routing path and a few attempts directed to "any PR who can route this
packet toward the destination". The later is known as alternate
routing and becomes most useful when you've lost connectivity with
your neighbor (perhaps its driven under a bridge or behind a
building).

Because of the uncertainty of successful retransmission, PR nets count
on a reliable end-to-end protocol as well. One way to look at your
question is that PRnets, with per hop acknowledgements, approach the
non-lossy nets in probability of successful transmissions. I agree
with your thought that due to "congestion => dropped packets"
behavior, simply using hop by hop acknowledgements seems fruitless.

Jil



This archive was generated by hypermail 2.0b3 on Thu Mar 09 2000 - 14:39:35 GMT