[Babel-users] battling with babel and route changes

Dave Taht dave.taht at gmail.com
Wed Jun 22 14:46:23 UTC 2011


2011/6/22 Juliusz Chroboczek <jch at pps.jussieu.fr>:
> [Copying to babel-users, by Dave's permission]
>
> Dave, sorry if you already have, but could you double-check that you're
> not NAT-ing or using a stateful firewall in your tests?

No. The results I get are without nat or firewalls, in a fairly
tightly controlled situation
in two labs described in my previous mail. We've stomped on a few
critical Linux bugs in the last month however that I am only now
preparing to test against.

There IS (now) a great deal of packet classification happening re
"saving the mice"
which I'll get to in a subsequent mail... but the means are currently
so computationally intensive that it's suitable only for research at
the moment. [1]


>> but as for the issue of the route change problem perhaps it would be
>> kind of cool to always have two routes installed of 1 different metric
>> to eliminate the window, and changes like 'unreachable' messages move
>> faster than babel can.
>
> Babel does maintain alternate routes, and will switch immediately (as in
> within microseconds) when a route is detected to be unreachable.  What
> takes time is detecting when a neighbour goes down -- it takes up to
> three hello intervals (losing a single hello should not trigger
> a routing change).

The layer below babel is even more responsive to changes, particularly
in the case of wired.

What would keeping two kernel routes always available of different
metrics (priorities being the correct name for it in the kernel level)
do to babel's convergence and loop detection?

It would simplify the the route change code mentioned in the previous
mail to already have two routes present.

>
>> And to have a "I'm about to move" announcement...
>
> We already have that -- just send an announce with a high (but finite)
> metric, which causes other nodes to switch to their alternate routes.
> A student of mine (Marko Mijatović) is currently working on a version of
> babeld that uses this kind of tricks when power goes down.

Very cool. Is this what happens when link state is detected on a
switch from wireless to wired?

1: https://github.com/dtaht/Diffserv and
    http://www.bufferbloat.net/projects/bloat/wiki/RFC_Improving_DSCP_support_in_Linux

-- 
Dave Täht
SKYPE: davetaht
US Tel: 1-239-829-5608
http://the-edge.blogspot.com



More information about the Babel-users mailing list