The TCP Large Windows WG met for half a day to discuss the two
proposals (RFCs 1072 and 1106) for improving TCP for large
delay-bandwidth paths. During the meeting two key issues were
raised.
The group determined that a key problem was how large to permit
the window to be. A larger window makes it easier to consume
the 32-bit sequence quickly. An example may help here. If one
permits a window of 2^30 bytes, then in each round-trip time,
one quarter of the sequence space can be consumed, and in four
RTTs, the sequence space will recycle. However, a TCP cannot
cycle the sequence space until it is sure the TTL of prior
segments has expired (the forbidden zone problem). So, we were
faced with choosing window sizes, that at anticipated speeds,
didn't cause the sequence space to roll over in less than the
anticipated TTL. In the end, the group was uncomfortable with
this problem and has asked Van Jacobson and Bob Braden (both of
whom have looked at this issue in more detail) to attend the
next meeting.
Another issue was whether we preferred to use options in every
segment to expand the window, or preferred to find a way that
didn't cause implementations to do expensive option handling.
The consensus was to avoid option handling (which meant we
preferred the rfc 1072 approach). Some discussion was given to
generating a larger TCP header, but this conversation foundered
when we checked the TCP header and found it lacked a version
number.
The group did not have time to consider another interesting
proposal (passed on from the IETF Hosts group) to allow text
error messages in RST segments.
ATTENDEES
Borman, Dave McCloghrie, Keith
Elz, Robert McKenney, Paul
Fox, Richard Miller, Dave
Galvin, James M. Solensky, Frank
Hedrick, Charles St.Johns, Mike
Karels, Mike Yasaki, Brian
Love, Paul
