Cost Model

 

Instead of examining each of the earlier examples in more detail, we will now concentrate on a common model and study the events and associated costs on a given node. Figure 2.1 depicts the principle graphically. The arrival of one or more messages triggers a small amount of computation that results in zero or more messages being sent.

  
Figure 2.1: Common Events: A node receives one or more messages, does some processing, and then, potentially, sends one or more messages.

We now consider the events that take place on a node from the time a message arrives until the handler completes and normal operation on the node is resumed. There are three cases.

1. Architectures that require supervisor (kernel mode) privileges to access the network interface.
2. Architectures that map the network interface into the user's address space.
3. Architectures that dedicate a general purpose CPU to the handling of communication.

For each architecture we consider the cost of a handler running in user mode, including the necessary context switches, and compare it with the case where the handler is executed in the kernel itself.

 
• Restricted Access to Network Interface
  • Handler at User Level
  • Handler in Kernel
• Mapped Network Interface
  • Handler at User Level
  • Handler in Kernel
• General Purpose CPU as a Message Coprocessor
  • Handler at User Level
  • Handler in Kernel
• Analysis