Juliusz, it's a pleasure to meet you. I've seen your name quite
often in the async/await world. Admittedly, usually in the detailed
"how things work" part - while I tend to be on the "teaching how to
use an implementation" side of things.

> > Dijkstra's algorithm remains a very natural approach to mapping a
> > graph
> I'm not sure what that means.  Dijkstra's is a shortest path algorithm,
> it's not in the business of mapping.  I guess the author meant that
> representing a graph as an adjacency list (the LSDB) is natural, which is
> certainly true, but in no way specific to OSPF.

Absolutely. Most of my development background is in game development,
I also do a lot of GIS. In both fields, Dijkstra's algorithm - and its adaptations
(A*, weighted flow maps, etc.) refer to mapping in the spatial sense; and converting
a map to a node graph (whether grid, waypoint, etc.) and then working with
cost-based adjacency (not raw adjacency) is a very natural way to
resolve the issue of "how do I get from X to Y" on a map. It's in no way
specific to OSPF (although specific adjacency cost specification was
one of many reasons OSPF outperforms RIP).

OSPF is where it is now because it's "good enough (for now)" and just
about everything supports it. Sure, an implementation that spits out bad
LSAs is going to break everything - you're going to get some pretty nasty
results from sending out broken destination-distance-vector data, too.
Garbage-in, garbage-out is one of the few truly universal rules! I agree,
though - I wouldn't hand out large-scale OSPF administration to the new
guy (although "here's the standard router config, plug in the numbers for
the locally attached networks here" does work).

I'd love to see good support for dynamic capacity analysis, unequal
cost multipath and similar. Babel looks very promising, but the chicken-egg
problem is very real; I can't put it to use until it's widely available, but
it won't become widely available until enough people put it to use. (It
seems like wireless vendors are busy trying to reinvent it at layer 2 with
proprietary meshing that doesn't talk to other proprietary meshing; ugh)




On Sat, Oct 29, 2022 at 4:15 AM Juliusz Chroboczek <jch@irif.fr> wrote:
> our toasts to the builders of Notre-Dame.

...which then burnt down :-/

> Dijkstra's algorithm remains a very natural approach to mapping a
> graph

I'm not sure what that means.  Dijkstra's is a shortest path algorithm,
it's not in the business of mapping.  I guess the author meant that
representing a graph as an adjacency list (the LSDB) is natural, which is
certainly true, but in no way specific to OSPF.

> I don't suppose you have ever had any ideas to how to improve things?

Modern OSPF and IS-IS have pretty much reached a local optimum: all the
low-hanging fruit has been picked, I doubt there's much that can still be
done to improve them without a complete redesign.  Well-implemented OSPF
and IS-IS work beautifully in a well-administered network, any other
protocol is going to converge slower and give less visibility into the
network.

On the other hand, OSPF is extremely fragile in the presence of bad
implementation.  If two routers have the same id, OSPF is going to create
routing pathologies.  If a router corrupts its LSDB (for example due to
bad RAM), OSPF will create routing pathologies which will only go away
once the faulty LSA expires (30 minutes worst case).  If a router runs out
of memory for its LSDB, it needs to stop participating in the protocol,
lest it cause routing pathologies (IS-IS has the overload bit to deal with
this case, which causes the router to become a stub router).  Compare this
with distance vector, where a corrupt routing table entry will only
interfere with the traffic to that particular destination, and where it is
perfectly correct to run with a partial routing table.

OSPF also requires a skilled administrator.  Splitting a network into
areas without causing suboptimal routing takes significant skill, route
filtering can only happen on area boundaries, and there are multiple
different ways of redistributing routes into OSPF (external LSAs).

In my opinion, you want to be running OSPF in parts of your network that
are implemented with reliable gear and are managed by a competent
administrator, but you'll prefer a modern distance-vector protocol
(somebody mentioned Babel) where the hardware is cheap and the
administator is busy with other things.  Fortunately, due to the
flexibility of route redistribution in distance-vector protocols, you can
do both: a stable backbone using OSPF, and unadministered Babel bits at
the edges.

-- Juliusz