Simplicity was a major consideration in Ethernet design. Computers on an Ethernet LAN are connected to each other directly by a passive communication medium, the Ether. There are no intermediate stages that handle message routing; the sender puts its packet directly on the Ether, thus broadcasting it to all potential receivers that are hooked up to the LAN. Flow control is distributed uniformly across the network. Each node determines when it is safe to send its packet, i.e. when no other node is currently transmitting. The receiver snoops the Ether for packets addressed to it, and acknowledges receipt when transmission is successful.
The high reliability of this design is due to the fact that there is no single point of failure. All communications occur directly through a passive medium, which means that there are no complex routers that can fail and impair, or even halt, packet transmission. The only complex pieces of hardware on an Ethernet LAN are the nodes themselves; in theory a failure of any number of them should not affect other nodes in any way, since every node is responsible for handling all of its own interactions with the rest of the network. The absence of expensive switching hardware ensures that the cost of the network itself is negligible compared to the cost of the nodes. Finally, local flexibility of an Ethernet is due to the fact that nodes can be reconfigured (and their number varied) simply by plugging them into the Ether at any convenient point.
Ethernet possesses serious disadvantages as well. The statement made earlier about the absence of single point failure is not completely true. If a malfunctioning node fails to disconnect itself from the Ether, the whole network goes down. However careful the designers of the hardware interface are, there is no 100% guarantee against such occurrences. But the principal problem with the Ethernet has to do with performance and scalability, not reliability. A single shared medium of transmission limits the total bandwidth of an Ethernet by serializing all communications. As a result, the throughput can be increased only by substituting a higher-throughput Ether. Scalability is also a major problem: if too many nodes are hooked up to a single Ether, they will increasingly interfere with each other. More conflicts mean wasted retransmission tries, and performance degrades substantially.
In order to create a network with substantially better performance and scalability, Autonet designers had to change both the routing strategy and the underlying hardware. Message routing is handled by active packet switches, connected by point-to-point links. Every switch must have buffers to store the queue of incoming messages. Special software is used to select the best route for each packet from the available alternatives.
This approach delivers exactly those advantages which Autonet was designed to provide. Point to point routing and multiple pathways from node to node enable messages to go through the network concurrently. Overall bandwidth can be increased to many times that of a single link by adding new packet switches, or new links to existing switches. Thus, the throughput of the network is not limited by the capacity of a single link, as is the case with Ethernet. Packet switches can use pushback to regulate the flow of data, reducing the number of discarded packets to near zero under normal circumstances. That results in fewer retransmissions, further increasing useful throughput. Autonet also scales better than Ethernet, since shared hardware (packet switches) can be scaled at will.
The increase in performance is achieved at the cost of simplicity and reliability, i.e. exactly the areas where Ethernet excels. Packet switches are much more prone to failure than a simple wire; even with multiple pathways, problems with only a few of the switches can significantly slow down the network traffic, or even disconnect parts of the network from each other. Dealing with this problem requires redundant hardware and special control software, which is capable of reconfiguring the network on the fly. Software complexity is further increased by the need for avoiding potential deadlocks, and periodically probing adjacent links for health. Both the hardware and the software cost money to develop (or buy) and maintain. Therefore, Autonet is preferable to Ethernet only in high-end networks, where good performance is essential.