renglish@hpirs.HP.COM (Robert English) (04/08/89)
I'm looking for a way to allow a network server to move between processors connected by a LAN. At the IP level, this can be done by using unsolicited ARP messages to inform other processors that the server has moved. This approach, however, has some problems. First, the unreliability of message delivery leaves the system open to the "black hole" phenomenon, where remote sites keep the old station address in their caches and continue to send messages to the wrong processor. Second, not all IP sites handle unsolicited ARP messages correctly, so that even if the message were delivered to the remote site, it might not change its routing table appropriately. And third, not all LAN protocols are based on IP, so that the approach simply doesn't help with servers that use both IP-based traffic and non-IP-based traffic. A solution to this problem seemed obvious to me: Use the programmable multicast address capability on the card to allow a server's station address to move with it. Since the mapping between station address and IP address does not change, black holes are not a problem. Since the station address itself moves with the server, servers that use non-IP protocols can be supported as well. Everything seemed to work fine. But then I learn that the Host Requirements RFC explicitly forbids the use of physical network multicast addresses for standard IP messages, to the extent that hosts receiving standard IP packets from multicast sources are required to drop them. So long, obvious solution. Apparently, the multi-cast IP spec doesn't allow for this possibility, either. Multicast IP addresses are legal as destinations, but not as sources, so they wouldn't help solve my problem, either. I can see many reasons to be careful about using multicast station addresses in a LAN, but very few to forbid them. Suppose, for example, that multicast addresses were legal only if they were derivable from the IP address they were associated with. For example, a host with the IP address x could only use the multicast address 0x800....0|x, and all others would be illegal. In that case, since IP addresses must be unique on a single LAN, then the multicast station addresses for the hosts would have to be unique as well. Does anyone have a better solution to this problem? --bob-- renglish%hpda@sde.hp.com
renglish@hpirs.HP.COM (Robert English) (04/12/89)
> / hpirs:comp.protocols.tcp-ip / renglish@hpirs.HP.COM (Robert English) / > But then I learn that the Host Requirements RFC explicitly forbids the > use of physical network multicast addresses for standard IP messages, to > the extent that hosts receiving standard IP packets from multicast > sources are required to drop them. So long, obvious solution. Hello, obvious solution. It appears that I misread the Host Requirements RFC. It doesn't outlaw link-level multicast source addresses for IP packets. It only outlaws link-level broadcast source addresses for standard (non-multicast or -broadcast) IP addresses. My thanks to the individual who took the time to set me straight. --bob-- renglish%hpda@sde.hp.com