In message <[email protected]>
Juliusz Chroboczek writes:
> > Another topic comes to mind. The topic is the partitioned bridged
> > network.
>
> The first situation was informally described by R. Tomlinson as
> a "network partitioning" problem in which a particular host, H in
> network N, is reachable from one gateway attached to network N but not
> another, because network N has become partitioned into two or more
> pieces.
>
> [...]
>
> The third situation arises in connection with an advanced airborne
> packet radio application. It first emerged in conversations with
> Major L. Druffel of the DARPA/IPT office. In this case, long-range
> packet radios (200-300 miles) are installed in aircraft and on the
> ground at selected sites. The ground sites may or may not have
> connectivity with each other (e.g., through a wire network and
> gateways). While aircraft are aloft, they communicate with each other
> and the ground via packet radios. If we treat the ground packet radio
> networks as a single net (for internet addressing purposes) and
> include the airborne packet radios as a part of that net, then this
> creates the partitioned network problem which was raised by
> R. Tomlinson.
>
> -- Vint Cerf, IEN 110, 1979
>
> While I happen to be an advocate of advertising a /128 on the host, as you
> suggest, I'd like to point out that HNCP might finally solve this 40
> year-old issue. Draft-ietf-homenet-prefix-assignment is not very clear on
> this issue, but I think that it implies that the algorithm is re-run when
> a link is partitioned -- which causes renumbering and causes two prefixes
> to be assigned to the now partitioned link.
>
> Pierre, is that correct? Does it need clarifying in your draft?
>
> Interestingly enough, the memo cited above describes the issue of
> multihomed hosts losing TCP connections when one of the networks goes
> down -- an issue that MP-TCP is finally solving 40 years later.
>
> -- Juliusz
Juliusz,
If HNCP renumbers on a partition, then existing TCP connections would
drop.
Numbering the loopback avoids this. Also there is the solution used
in the 1990s by ISPs which you trimmed from my email, but requires
installing host routes.
Also consider what might happen with an intermittent link between
switches if the solution is to renumber on a subnet partition (which
should occur, but for the interface addresses only, if not using link
local).
My preferred solution is:
1) always number the loopback except for legacy hosts,
2) use link local interface addresses on a point to point interface
including Ethernet with two hosts except for legacy hosts,
3) number any broadcast interface using HNCP (or since this is my
preference, extend OSPF - no flames please).
Thanks for the historic reference.
Curtis
_______________________________________________
homenet mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/homenet
