LYNCH@USC-ISIB.ARPA (Dan Lynch) (12/13/85)
Dave, Your last line was a whopper: that new procurements have subnetting support written into them! That is a rathr drastic measure and only pushes off the day when gateways will have to get smart anyway. I would vote for putting the most intelligence (and hairy algorithms) in the fewest possible places. That says to make gateways smart and let the hosts play dumb. Let there be a big database in the sky somewhere (known only to gateways?) that knows all routes and connectivity gory details. Then the gateways should just act as a cache for that monster database and feed their clients (hosts) info on a demand basis. This principle harkens from the days of early radio design when someone called a halt to the (then) impending decision to make the "protocol" between transmitters and receivers well "balanced" by spreading the implementation costs equally. The observation was made that when things got going hot and heavy there would be millions of receivers and only thousands of transmitters, so put the hair in the transmitters and make the receivers as simple as possible. (I am indebted to Dave Boggs for this story.) So, I argue for making life as simple as possbile for the multitude of hosts and put the hair in the gateways. Dan -------
MILLS@USC-ISID.ARPA (12/13/85)
In response to your message sent 12 Dec 1985 13:12:49 PST Dan, I've never heard of Dave Bogg's story, but it probably would not apply to the non-consumer radio community today anyway. I'm not sure I understand what the gateways can do to ameliorate the the problem that (some) hosts do not understand subnets, other than in the special, but common, case of Ethernets, where Noel's suggestion might in fact be incorporated in the present gateways without too much trouble. His technique, which I have called "promiscuous ARP" is used in our gateways as well. As for advanced models, you may wish to see the document NEWMOD.DOC in the MILLS directory on ISID, which describes one possible scenario supporting mobile hosts, multiple-destination, multiple-path routing and serves ice cream for dessert. I suggest further discussion on this topic be switched to the tcp-ip repeater or direct. Dave -------
mogul@SU-NAVAJO.ARPA (Jeff Mogul) (12/13/85)
I disagree that requiring RFC950-style subnet support is a bad idea. If implemented reasonably, using address masks, it insulates a host entirely from the addressing format (i.e., subnets? and if so, what bits specify the subnet?) So, rather than moving the smarts into the hosts, RFC950 goes in the opposite direction - put a simple mechanism into each host that allows it to offload all the smarts onto the gateways. Of course, even this much change is probably too much for some of our vendors. But the internet is too big to pretend that subnets are unnecessary. -Jeff P.S.: I believe Noel Chiappa has consistently advocated the position that non-gateway hosts should not even have routing tables; just a large enough redirect table (with host redirects only) and the address of one or more neighbor gateways.
JNC@MIT-XX.ARPA ("J. Noel Chiappa") (12/13/85)
Dan, the suggested modification to the host IP layer (to support subnetting), and other changes in gateways that the Gateway Committee are contemplating, are supposed to make the host IP layer simpler, not more complicated. The whole idea is precisely to put all the smarts in the gateway, with only the simplest possible cache in the host, and to attempt insulate the host IP layers from any further changes to the basic system architecture. Once all hosts use the mask technique, and get (and use) only per-host Redirects, it will be possible for the gateways to play arbitrary games with the layout of the system without the host IP layers noticing. The sooner this change is installed everywhere the better. Noel -------