Paper review: Ethernet: Distributed Packet Switching for Local Computer Networks

Reviewer: Hanlin D. Qian

This paper investigates and addresses the problem of how to connect a group of local computers in a cost-effective way. The paper makes the assumption and argues that any type of centralized control in the lower level network design would be expensive, complex, and inefficient. Therefore the authors propose a networkof computers where control is completely distributed among the stations in the network. Packet transmissions in such a system is coordinated by statistical arbitration, which means that packets have a high probability of arriving at its destination but is not guaranteed to get there. There are several advantages to this approach:

1. The network setup is cheap and inexpensive; the stations on the network are clearly more expensive than the network itself.
2. The network is scalable; it's easy to add stations to the network without modifying the overall scheme of the network.
3. If a station breaks down, the rest of the network can continue to operate because there is no central control.

The design of the network is based on the principle of passive broadcast packet switching. Packets are sent out into the Ether from individual stations so that every other stations can receive them. Thus packet collision can easily result and corrupt packets. Other problems such as interference, impulse noise, inactive receiver at a packet's destination, and an intentional discard can all cause packet losses.

This paper proposes the method of random back-off and retransmission to address these problems. This solution simplifies the network because it places the burden of verifying the arrival of the packets at the end points. The paper also proposes the following mechanisms and heuristics to address and improve the overall performance of the Ethernet.

1. The transceivers use the time interval between the time when a packet is detected and the time when a collision is detected to calculate a random backoff time period before retransmitting the packet.
2. Carrier detection allows a transceiver to delay the transmission of a packet if another packet is passing the transceiver at the time of transmission. This mechanisms reduces collisions.
3. Interference detection can detect the corruption of packet within one round trip time, long before an acknowledgement timeout.
4. A checksum is placed at the end of the packet to verify whether an arrived packet is correct.
5. Truncated packets due to collisoin can be filtered out before it reaches it destination and therefore reduces the network load.
6. Momentary jams in the Ether can inform all stations that a collision has occurred. Thus stations can then abort, backoff for a random amount of time, and retransmit a packet.

The paper designs a detailed model of this Ethernet system to test the design and hypotheses. The experimentation did take place with a moderate but reasonable set of constraints. A system of 256 stations and 1 km of cables was contructed. Protocols such as EFTP were also developed to run tests. The results of these tests show that the assumptions made in the design of the Ethernet are good ones, and the Ethernet can run efficiently.

I give this paper a high rating of 4 because it not only designed a model of the Ethernet but also backed it up with concrete tests, experiments, and mathematical models. However, I do have to add that the experiments described on the paper are neithe exhaustive nor extensive. For the limitation of the scope of the paper, only one setup environment has been tested. Further research and testing needs to be done to experiment with networks of different sizes and loads.