28 Mar
2021
28 Mar
'21
3:48 p.m.
On Mar 27, 2021, at 2:54 PM, Johnny Billquist <bqt
at softjar.se> wrote:
Paul and me have lamented over this many times.
DDCMP assumes that the underlying transport mechanism is reliable. Both in terms of the
ordering of data, the reliability of getting all data across, and no duplication, and so
on...
UDP does basically not fulfill any of these requirements, making UDP a very unsuitable
transport for DDCMP links.
No, you're confusing DDCMP and Multinet.
Multinet isn't a datalink protocol; it's merely a wrapper for the routing layer
packets. In particular, it wraps routing layer point-to-point sublayer packets. THAT
assumes a reliable data link.
DDCMP is en entirey different beast. It provides that reliable data link, and it sits on
top of an unreliable physical layer. It handles packet loss mostly because it handles bit
error, which causes packet CRC error and is handled by timeout and retransmit
("ARQ").
...
But as far as DDCMP goes, you should essentially not use UDP as a transport. That is just
bad. Use TCP instead. And then you're good.
No. It's certainly ok to use DDCMP over TCP -- except that it has to deal with the
oddity of doing both connect and listen concurrently. But UDP works fine.
Johnny
On 2021-03-27 11:40, Keith Halewood wrote:
> Hi,
> I might have posted this to just Paul and Johnny but it?s probably good for a bit of
general discussion and it might enlighten me because I often have a lot of difficulty in
separating the layers and functionality around tunnels of various types, carrying one
protocol on top of another.
> I use Paul?s excellent PyDECnet and about half the circuits I have connecting to
others consist of DDCMP running over UDP. I feel as though there?s something missing but
that might be misunderstanding. A DDCMP packet is encapsulated in a UDP one and sent. The
receiver gets it or doesn?t because that?s the nature of UDP. I?m discovering it?s often
the latter. A dropped HELLO or its response brings a circuit down. This may explain why
there?s a certain amount of flapping between PyDECnet?s DDCMP over UDP circuits. I notice
it a lot between area 31 and me but but much less so with others.
I need trace logs for that, because that isn't supposed to happen.
More precisely, DDCMP definitely (this is tested) handles random packet loss. It will
retry indefinitely, as do other DDCMP implementations. But the routing layer has its own
"hello" protocol which has a limited tolerance for blocked communication, so if
there is a failure to communicate that lasts too long (the "listen timer") then
you will get a link down. For DDCMP, a fairly long hello timer (60 seconds or so) is
recommended.
If you see something else, in particular anything that suggests packet loss isn't
handled transparently, please let me see the trace data.
paul