Date:        Mon, 02 Feb 1998 15:26:20 -0800
    From:        "Kevin M. Lahey" <kml@nas.nasa.gov>
    Message-ID:  <199802022326.PAA19554@gecko.nas.nasa.gov>

  | >	(1) No Black Hole Detection
  | There aren't alot of obvious ways to do this, but we'll try.

Aside from the MS NT way described (which seems to me to be akin to
cheating, and will be impossible for IPv6) would it not be reasonable
when no ACKs are being returned (ACKks imply no black hole of course)
send a baby packet - window probe or keepalive style - and see if that
is ack'd.   Make it so small that it cannot require fragmentation (<80 bytes).
If that gets an ACK, then it is very likely that big packets aren't
getting through, so halve the anticipated mtu (or something similar) and
continue.   Not only does this catch the case of improper DF processing,
or filtered ICMP, it will also allow communicatins to work over links
where (due to clocking problems or whatever) essentially all long packets
manage to acquire crc errors.

  | We can always crank down the MTU as we detect timed out packets;

That's probably overkill without some better evidence.

  | we can't always depend on maintaining the same path during
  | the course of a connection, 

Of course.   You must expect the PMTU to vary over time (which includes
occasionally sending bigger packets to see if the PMTU has risen).

kre