A collision domain is a networking term that refers to one or more physical network segments joined by a repeater hub. Cisco refers to this as the 'media access domain', that is, a collision domain consists of all the devices connected to the network that must share the network's physical connections and bandwidth. Transmissions from two or more sources (computers, routers etc.) within this joined network can occur at the same time with each host's message obliterating the other host's message on the wire. This interference is referred to as a collision. The Ethernet network protocol was designed to expect collisions and has a collision avoidance function, but collisions reduce overall throughput on the network. Ethernet networks using hubs, bridges or repeaters are particularly likely to experience collisions. Data transmitted by any host into a hub is flooded out all other ports on the hub. Hubs, bridges and repeaters increase the collision problem by bridging together collision domains by retransmitting all recieved traffic to all ports.
The size of a collision domain is equal to the the time it takes one frame of data to travel to the far side of the network and back. This is directly impacted by the propagation delay of the medium used. Fiber optic networks can span physically larger distances, copper networks span smaller distances than fiber networks. Ethernet frames are of variable size, but the smallest frame is 512 bits, so for Ethernet, the collision domain can't be any larger than the time it takes the 512 bits to travel round trip between the two most distant points in the network (at least 920 meters for copper-based 100Mbps Ethernet). If the distances are larger, the two hosts will not recognize a collision has happened and will hapilly continue to transmit, resulting in a communications failure at the link layer.
There is a second problem with a collision domain that encompasses the entire set of hosts within an enterprise. As the volume of network traffic increases the chances of a collision occurring increase. Each collision slows down the transmission process as both hosts stop, wait a random period of time and then retransmit, hopefully without coliding with yet other transmissions. As the number of collisions increases, the overall throughput of the network segments decreases until none of the sources can transmit--everyone is trying to retransmit or avoid a collision, so no data can make it through the network segments comprising the collision domain.
To reduce a collision domain, physically segment the network by reducing the number of hosts that share a network and create more than one physical network. Segmenting the network physically creates separate networks that cannot communicate with each other. This is usually not acceptable. The second choice is to logically segment the network by replacing your hubs and bridges with switches or a router. Hubs and bridges are growing increasingly rare but they still exist and can create significant network communication and performance issues.
Switches resolve the collision domain problem by splitting the collision domain into fewer sets of hosts where fewer hosts can see the traffic from other hosts. Switches begin by flooding all frames out all interfaces, but as they learn the MAC addresses of the hosts connected to the switch, they begin switching frames from port to port instead of flooding all ports with the frames.
Switches still perform flooding of certain kinds of network traffic such as ARP, RARP and broadcast packets from the network layer