From
http://www.sangoma.com/fguide.htm
Consider now two separate Ethernet networks that are joined by a PC, C, acting as an IP router (for instance, if you have two Ethernet segments on your server).
http://www.sangoma.com/images/tcpip_image002.gif
Device C is acting as a router between these two networks. A router is a device that chooses different paths for the network packets, based on the addressing of the IP frame it is handling. Different routes connect to different networks. The router will have more than one address as each route is part of a different network.
Since there are two separate Ethernet segments, each network has its own Class C network number. This is necessary because the router must know which network interface to use to reach a specific node, and each interface is assigned a network number. If A wants to send a packet to E, it must first send it to C who can then forward the packet to E. This is accomplished by having A use C's Ethernet address, but E's IP address. C will receive a packet destined to E and will then forward it using E's Ethernet address. These Ethernet addresses are obtained using ARP as described earlier.
If E was assigned the same network number as A, 200.1.2, A would then try to reach E in the same way it reached C in the previous example - by sending an ARP request and hoping for a reply. However, because E is on a different physical wire, it will never see the ARP request and so the packet cannot be delivered. By specifying that E is on a different network, the IP module in A will know that E cannot be reached without having it forwarded by some node on the same network as A.