Why does ROS have nodes?

edit

From http://wiki.ros.org/Nodes

The use of nodes in ROS provides several benefits to the overall system. There is additional fault tolerance as crashes are isolated to individual nodes. Code complexity is reduced in comparison to monolithic systems. Implementation details are also well hidden as the nodes expose a minimal API to the rest of the graph and alternate implementations, even in other programming languages, can easily be substituted.

Separate nodes gives better fault tolerance and debugging; it's much easier to see what's broken when a single node is crashing, and the OS prevents nodes from overwriting the memory of other nodes.

Using a pub-sub middleware like ROS also hides the thread/process synchronization that happens in any large, multithreaded system and provides a more obvious way to distribute the processing load across many cores.

Obviously there's a performance hit for serializing and deserializing messages, but for small messages this is negligible. If you're passing around large messages such as images or point clouds, the performance gain from using nodelets outweighs their complexity.

For more best practices, have a look at the ROS Patterns and Best Practices section of the wiki.