From: Howard C. Berkowitz (hcb@gettcomm.com)
Date: Wed Jun 16 2004 - 22:58:53 GMT-3
At 9:40 PM -0400 6/16/04, Calton, Doug wrote:
>A lot of the practice labs I have used include redistribution with
>specifications for being loop free and using optimal paths. As I work
>on these labs, it seems to me that "optimal" is VERY subjective,
When I write a practice lab, I don't such say "optimal." I tell you a
goal of optimization, which, most frequently, is stability versus
maximum information. Maximizing information, in real-world networks,
may still not give you optimal paths, although it will be more
likely. Maximizing information is also more likely to overload the
routing control plane, or induce oscillations into the routing system.
If I want you to have a route with QoS, I'll tell you either to
statically route it or set up an RSVP path.
>particularly in some of the nastier labs with a bazillion overlapping
>and alternate connections between routers. E.g., with loopback I/Fs,
>the optimal path to them will depend on just where you are coming from
>in the network, so what I remotely view as an optimal path to the loop
>may not agree with what the host router itself sees as the optimal path,
>leading to disjoint paths or even feeback/loops.
>I guess I have come to the conclusion that optimization can occur only
>within the IGP domain, if anywhere. Once you begin redistributing, all
>you can do is rough things in a reasonable fashion, and there is no
>clear "right" answer.
Speaking strictly of routing protocols without traffic awareness, I
agree completely with you. A fundamental problem of the lab is that
it overemphasizes mutual redistribution, where real-world large
enterprise network are generally hierarchical with local
optimization. By local optimization, I suggest that various regions
can each run an IGP domain and optimize within that domain. The
domains are linked by a backbone of backbones, usually BGP, that
first and foremost defines reachability rather than optimum paths.
Hierarchy is such that you advertise routes upward, but generally get
default downward. Within a level, you can take the cheapest local
path to a default.
Getting into the global Internet, researchers have held for a very
long time (in Internet years) that it is impossible that the global
routing table will ever completely converge. We use VPNs and traffic
engineering to get tuned paths.
>Any thoughts on how solutions are evaluated based on this, or other
>thoughts?
I don't really know what to say. The lab is such an artificial (and
small) environment that any resemblance to real-world optimality
tends to be coincidental. Indeed, the emphasis on exercising obscure
commands often leads to suboptimality, whatever that may be.
This archive was generated by hypermail 2.1.4 : Sat Jul 03 2004 - 19:40:42 GMT-3