6.033 Rod Ferguson Paper #3 February 28th, 2000
Provisions for Scalability Sacrifice Reliability
Mechanisms used to make Ethernet scalable led to various sacrifices from the original design. Metcalfe and Boggs saw scalability as an important aspect of the Ethernet system. They identified the limitations to the number of connected nodes as the length of Ether over which the required signals could be carried, the finite bandwidth, and finite address size. Their respective solutions were packet repeaters, filters, and gateways. Packet repeaters and filters sacrificed fault isolation and packet gateways cause the system to suffer from added complexity.
Signal cover was extended by the introduction of the packet repeater. The packet repeater joined together multiple segments of Ether, and thus expanded the number of connected nodes. The packet repeater would pass packets unconditionally between the Ether segments it connected. However, due to interference issues, there could only be one packet repeater between any two segments. Failure of a packet repeater would therefore sever all communication between nodes on different segments that it once connected. This is a direct contradiction of their original desire to have "the failure of an active element...tend to affect the communications of only a single station."
The packet filter, a variation on the packet repeater, was used to extend traffic cover. The packet filter would only pass packets to another Ether segment on the condition that the destination station was located on that segment. While this restricted traffic to Ether segments appropriate to each packet, it otherwise inherited all of the shortcomings of the packet repeater.
Metcalfe and Boggs offered the mechanism of packet gateways to extend the address cover of the Ethernet. A packet must address a gateway as an intermediary in order for the gateway to pass the packet on to another Ether segment. This specific addressing of the gateway allows there to be multiple gateways connecting two segments without interference problems. Thus, if there is a failure in a particular packet gateway, the sender can retransmit the packet through a different gateway, and fault isolation is preserved. The simplicity of the orginal design, where the sender only needed to know the address of its intended receiver, is lost with the use of gateways. Gateways cause nodes to have multiple addresses, and the sender must choose one of these depending on its location with respect to the location of its intended receiver. This added complexity reduces reliability as it is more prone to error.
Ethernet was designed to be able to grow smoothly, allow communication between stations independent of their locations, and maintain fault isolation. On a small scale, ie. with one Ether segment, these goals are realized, but attempts to allow growth beyond a single Ether segment, have led to undesirable sacrifices.