Imagine asking a computer to read a large file. While the disk supplies the next piece of data, the processor may have little to calculate. Work is I/O-bound when faster input or output would make it finish sooner. The bottleneck is the device or service it is waiting for, not the speed of the processor. section 6.3
Waiting can take several forms:
- A search through a huge file can be limited by how quickly the disk supplies its contents. section 6.3
- A web server or database may spend much of its time on network or disk I/O. section 6.3
- A graphical application may wait for a person to use the keyboard or mouse, even when it never opens a file. section 6.3
The distinction matters when choosing a performance improvement. Multiplying enormous matrices is CPU-bound: computation is the limiting work. Copying a large stream from a network to a disk can instead leave cores waiting for data transfer. Adding processor cores alone will not remove that I/O bottleneck. section 6.4
Keeping an arcade game responsive
Picture a one-core arcade game with three jobs: get_user_input() waits for the controller, compute_game_world() calculates what happens next, and render_next_screen() sends the result to the display. Input and rendering are I/O-bound; world computation is CPU-bound. When the player has not pressed a button, the processor can work on another ready task instead of spending that time waiting. section 6.5 section 6.8
An operating system can park a task blocked on I/O, run a ready task, and make the parked task ready again when I/O completes. This is why multitasking can help I/O-heavy workloads even on one core: it uses otherwise idle processor time, rather than making the device faster. Switching tasks has a cost, however; creating more tasks is not automatically an improvement. section 10.10 section 6.14
