> ...
> > The "Ongoing TCP Research Related to Satellites" draft suggests:
> > "Hoe [Hoe96] proposes using the packet-pai algorithmr [Kes91] to determine =
> > a more
> > appropriate value for ssthresh."
I've posted this before but here it goes again. In general the packet pair
algorithm is very hard to apply to TCP in its startup phase. The reason is that
because of the receiver delaying the ACKs, two packets sent back to back (as
required by packet pair) wouldn't necessarily cause the receiver to respond
with an ACK for each immediately. In fact the ACK spacing can be such that you
end up grossly *underestimating* your bandwidth.
Hoe suggests measuring the bandwidth using 3 closely spaced ACKs. As a TCP
receiver sends an ACK for every 2 segments, 3 closely spaced ACKs would only
be generated if the sender sends 6 closely spaced segments. But in slow-start,
the sender can send 6 closely spaced segments only upon receiving 3 closely
spaced ACKs - this becomes a chicken and egg problem. Thus the only way
by which the sender can send 6 closely spaced segments is to start with an
initial window of 6 segments - something that is considered inappropriate.
- Mohit Aron
[email protected]
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:43 EST