JNC@MIT-XX.ARPA ("J. Noel Chiappa") (12/04/85)
Actually, I'm a little bit puzzled by this space issue. In the CGW, a route takes up 14 bytes while an EGP neighbour takes up about 50. Admittedly, some of the fields in the EGP block aren't use simultaneously, and could be overlapped, but it's still not small. What I find curious is that the core gateways will apparently happily take many neighbours, but don't have room for routing entries. Why are the routing entries in the BBN gateways so large? Can someone from BBN explain what's going on here? Noel -------
brescia@BBNCCV.ARPA (Mike Brescia) (12/05/85)
The BBN gateways current version does not use memory mapping in the lsi11. (There are even some of the machines which are 11/02 without mapping hardware.) The tradeoff in memory is between packet buffers and routing table entries. Currently, the buffer capacity at some sites is just enough to absorb a bunch of packets from an ethernet (fast) sending out to an arpanet (slow) for a single ftp connection. In those cases, losing one buffer causes the probability of packet dropping to increase dramatically. Just ask the people at ISI. Memory mapping is included in a new version of software which is now running on a machine between bbnnet and arpanet and a local ethernet (which is to say that it is beyond debugging and into testing). In a couple of weeks, it should be released to some sites which have memory mapping hardware (11/23 processors). The arpanet-milnet gateways are being placed under configuration management, and should be ready for release with memory mapping in a couple of months. Memory mapping will allow extra memory to be used for buffering and allow more networks to be listed. Mike Brescia
LYNCH@USC-ISIB.ARPA (Dan Lynch) (12/05/85)
When I saw the first message on this topic I thought the issue was one of an outdated algorithm for routing table maintenance and not one of just increasing the table size to get over the current hump. Since some of the gateways will never have much memory (when is there ever "enough"?) it would appear that a name server model for gateways is in order. Hosts already have to deal with this issue (if they are playing it "honestly"). Dan -------
MILLS@USC-ISID.ARPA (12/05/85)
In response to the message sent 04 Dec 85 16:35:32 EST (Wed) from brescia@BBNCCV.ARPA Mike, While more memory might provide a few more nets, my comment on packet-size limitations presumably remains valid. Last we talked, I think something like 130-odd networks was expected to break, depending upon the exact mix of class A/B/C nets. I got a taste today myself, when our beloved gateway winced and we had to recompete the table entries. Do not bury the port-expander nets just yet... Dave -------