TCP/IP allows us to create very large internetworks by connecting individual networks together. When we send data between devices on different networks, the data is routed between networks using the Internet Protocol. This permits us to view even a huge internetwork such as the global Internet as if it were just a single large network. All devices on the internetwork are considered to be virtually connected at layer three, since the process of routing lets any device talk to any other one.

However, there is no way for devices on distant networks to communicate directly. The internetwork communication at layer three really consists of a number of steps, called hops, that carry the data from its source to destination. Each hop in a route requires that data be sent between a pair of hardware devices, and each transmission must use layer two hardware addresses. Since TCP/IP uses layer three addresses, this means each hop requires that we translate the IP address of the target of the hop to a hardware address. This is called address resolution; the reasons why it is needed and the methods used for it are explained in detail in the section on address resolution concepts.

In TCP/IP, address resolution functions are performed by the aptly-named Address Resolution Protocol (ARP). When a device needs to transmit to a device with a particular IP address, it can use ARP’s request/reply messaging protocol to find out which hardware device corresponds to that IP address. However, each such message exchange takes time and network bandwidth, so for efficiency, every device maintains an ARP cache, which is a table containing mappings between IP and hardware addresses. The ARP cache table can contain a combination of static cache entries that are manually inserted for frequently-accessed devices, and dynamic entries, which are entered automatically when a request/reply resolution is done, so the next time it is necessary to send to that device, the lookup process can be avoided.