UNIX System Administration Handbook - Evi Nemeth [226]
At layer 3, the network layer, routers are used. Routers transfer messages to the next hop according to the location of the final recipient, rather like looking at the message in a bottle to see who it’s really addressed to.
Hubs and concentrators
Hubs (which are also referred to as concentrators) are active devices that connect physical segments in UTP Ethernet networks. They require external power. Acting as a repeater, a hub retimes and reconstitutes Ethernet frames but does not interpret them; it has no idea where packets are going or what protocol they are using.
The two farthest points on the network must never be more than 4 hubs apart. Ethernet versions 1 and 2 specified at most 2 hubs in series per network. The IEEE 802.3 standard extended the limit to 4 for 10 Mb/s Ethernets. 100 Mb/s Ethernets allow 2 repeaters, and 1000BaseT Ethernets allow only 1. Exhibit C shows both a legal and an illegal configuration for a 10 Mb/s network.
Exhibit C Count the hubs
Hubs occasionally require attention from a system administrator, so they should not be kept in obscure or hard-to-reach locations. Power cycling usually allows them to recover from a wedged state.
Switches
Switches connect Ethernets at the data link layer (layer 2) of the ISO model. Their purpose is to join two different physical networks in a way that makes them seem like one big physical network. They do not require software, but rather receive, regenerate, and retransmit packets in hardware.3
Most switches use a dynamic learning algorithm. They notice which source addresses come from one port and which come from another. Packets are forwarded between ports only when necessary. At first all packets are forwarded, but in a few seconds the switch has learned the locations of most hosts and can be more selective.
Since not all packets are forwarded between networks, each segment of cable is less saturated with traffic than it would be if all machines were on the same cable. Since most communication tends to be localized, the increase in apparent bandwidth can be dramatic. And since the logical model of the network is not affected by a switch, there are few administrative consequences to installing one.
Switches can sometimes become confused if your network contains loops because packets from a single host appear to be on two (or more) ports of the switch. A single Ethernet cannot have loops, but as you connect several Ethernets together with routers and switches, the topology can include multiple paths to a host. Some switches can handle this situation by holding alternate routes in reserve in case the primary route goes down. They perform a pruning operation on the network they see until the remaining sections present only one path to each node on the network. Some switches can also handle duplicate links between the same two networks and route traffic in a round robin fashion.
Switches keep getting smarter as more functionality is built into their firmware. Some can be used to monitor security on the network. They record any foreign Ethernet addresses they see, thereby detecting and reporting newly connected machines. Since they operate at the Ethernet layer, switches are protocol independent and can handle any mix of high-level packet types (for example, IP, AppleTalk, or NetBEUI).
Switches must scan every packet to determine if it should be forwarded. Their performance is usually measured by both the packet scanning rate and the packet forwarding rate. Many vendors do not mention packet sizes in the performance figures they quote; therefore, actual performance may be less than advertised. Switches are a good but slightly expensive way to connect Ethernets.
Although Ethernet switching hardware is getting faster all the time, it is still not a reasonable technology for connecting more than a hundred hosts in a single logical segment. Problems such as “broadcast storms” often plague large switched networks,