[ih] IP over wireless [was: booting linux on a 4004]

Karl Auerbach karl at iwl.com
Wed Nov 27 12:37:58 PST 2024


On 11/27/24 11:39 AM, Brian E Carpenter via Internet-history wrote:

> As a member of the original IPng directorate, I believe that is 
> correct. But also, several of us had been severely burned by multicast 
> storms on large bridged Ethernets. So the implicit assumption was that 
> there wouldn't be any more of those, since moderately priced routers 
> were coming to market, and that it was therefore safe to use 
> link-layer multicast for neighbour discovery. And that is still true 
> on modestly sized Ethernets and Wi-Fi networks. Where it goes horribly 
> wrong is on *large* WiFis. I've done tests at an IETF meeting where at 
> most 10% of link-layer multicasts were delivered to a host sitting on 
> the same table as the sender. Don't put a thousand hosts on the same BSS.

Ah, "horribly wrong" - almost my middle name!

When we designed/built/operated the show networks for Interop in the 
latter half of the 1990s we heavily pushed IP (v4) multicast.

Our network was, for the time, very strongly provisioned over both 
copper and fiber (and some WiFi, which was new at the time) - we carved 
our 45/8 into on the order of 100+ /22 subnets.  We also reached out to 
venues across the host city.  We had a very large number of paths 
between pieces of the net - routing was complicated, including at least 
two external links to two external providers.

We had some fairly heavy IP multicast streams.  I had my Precept 
Software gear (designed and built by Steve Casner, Chia-Chee Kuan, 
myself, and a cast of very talented people.)  And we had Carl Malamud's 
Internet Geek radio.  And anything else the vendors wanted to throw onto 
the net.

We were using the classic IP any-to-any multicast; single source 
multicast had not come onto the scene.  We usually used PIM (sparse 
mode) for multicast routing.

We did have some serious issues with Sorcerer's Apprentice behavior, or 
at least infinite loops.  The largest was caused by a difference in 
interpretation about how to turn a directed IP broadcast into a MAC 
group or broadcast address.  The result was a fight between our Cisco 
and Wellfleet routers that caused infinite MAC level packet loops across 
most of our paths between subnets. Every load indicator LED in the 
convention center (and in the remote venues) turned red.  All other data 
stopped moving.  (And all eyes focused on me - because I was playing 
with multicast in the NOC when it happened, and because I have a 
tendency to be near things when they go awry, even if it's not my fault 
[although sometimes it was.])

         --karl--





More information about the Internet-history mailing list