On Fri, Jun 11, 2010 at 9:26 PM, Tony Li <[email protected]> wrote: >> 1. By a switch determining which port or ports to repeat a received packet >> on. > > Again, function 1 does not make the namespace topologically sensitive. It's > just using a flat identifier space and explicitly learning the location of > the identifier.
Actually, this all gets me to thinking... Instead of the ubiquitous Ethernet, what might a layer 2 network in which the MAC address was unambiguously a locator look like? 32 bits switch number. Assigned at the factory. 16 bits port number. The switch owns all 16 bits. NICs are not assigned any address. When a NIC first connects to a switch, the switch sends one address comprised of the switch number and the port number. When the switch connects to another switch, instead of assigning an address to the port, it announces its switch number and the other switch numbers it knows about as well as their distance metrics from itself. It likewise sends updates as new switches are learned and old carriers are lost. Everything else works as with Ethernet. 48 bit source and destination MAC comprised of the 32 bit switch ID and 16 bit port number transmitted by the host with each packet. Interesting beneficial characteristics of such a design: Smaller CAM. Need only keep track of switch numbers, not full MAC addresses. Only switches, not all hosts. Inherent loop mitigation. No need for spanning tree. Never a need to flood unicast packets to all ports. Can drop early if the destination switch is unknown. Optimal routing through a multiply connected switching mesh. Challenges with such a design: Simple repeaters that don't examine the MAC address not possible. Broadcast media difficult or impossible (e.g. wifi or 10base2) Must advise host network stack when a MAC is assigned. Stack may need to take further action, e.g. a broadcast or multicast packet advising of the new IP to MAC mapping. Picking a switch number when a host with multiple NICs or a host with multiple guest OSes needs to act as a switch. Deciding to emulate a switch if both ends of a wire are host NICs. Standard distance-vector protocol issues. Not really going anywhere with this or trying to make any points... just thinking out loud. Regards, Bill Herrin -- William D. Herrin ................ [email protected] [email protected] 3005 Crane Dr. ...................... Web: <http://bill.herrin.us/> Falls Church, VA 22042-3004 _______________________________________________ rrg mailing list [email protected] http://www.irtf.org/mailman/listinfo/rrg
