Subject: Re: Two incoming Internet links
To: None <port-i386@netbsd.org>
From: Kip Rugger <kbr@pangea.ca>
List: port-i386
Date: 08/13/2001 15:54:20
>> I have 256 Kbps from one ISP and 128 kbps from other, also I have one
>> firewall (ipf/nat) and three ethernet cards, how to do ipf/nat on one
>> ISP I know, but how to set up, that if one link is almost used and to
>> tranfer user to another ? Do I need to setup source routing ? And what
>> the is it ? :) Also I have heard, that in some cases firewall van be
>> configured that way: if from one ISP links is used up to 70%, to drop
>> clients to another ISP, or maybe I dream about it last night ? :)
>
>do you mean something like route switching? So that your network 
>users get connected to one ISP or the other depending on the 
>usage of the line? 
>
>You possibly could implement something to monitor traffic and 
>based on that findings in regular intervalls change the 
>defaultroute/gateway settings. That could probably work.


That sounds like good advice.

Look at the problem from a different viewpoint.  A large, multi-
national corporation with multiple ISP connections in diverse
regions, and with an internal network, is faced with the problem
of where to move a packet from the internal network to the public
network.  For example, for a packet originating in Los Angeles
destined for New York, should the packet leave the internal net
at LA, Denver, Buffalo Breath, New York, ...?

This generalizes to the problem of a network topology that is no
longer a spanning tree, that is, one in which ``loops'' are possible.
This occurred historically when the first internet ``backbones''
became multiply connected.

To handle such routing optimally, you require the routing metrics
for at least a subset of the *entire* internet.  This is what the
BGP (border gateway protocol) is about.  Of course, such routing
tables are themselves huge, and require corresponding resources.
ISP's do offer BGP (for a price), but usually to other ISP's or
telcos.

So, if you can't afford to do optimal routing via BGP, you are
left with sub-optimal routing, probably based on stochastic
measurements.  This is where you start your search for a solution.