Scheduling policy is a balancing act between competing goals. Modern scheduling policies
make tradeoffs between three primary goals: fairness, low latency and progress. Other goals
exist, but these three are often the most important. Fairness concerns how CPU cycles are
divided over some time scale (e.g. one second, one minute, one hour). A task's portion of
cycles over a given time period is called its CPU allocation. There is no quantitative definition
of fairness. A policy's fairness can be measured in how closely and at what time scale it matches
a desired allocation; the smaller the time scale the greater the perceived fairness. Scheduling
latency is how long a taks must wait before it is given control of the CPU. Latency is most
important for interactive tasks because high latencies result in frustrated users. Progess measures
the work a task can accomplish in a given time period. In the extreme case, called starvation,
a task may take no progress at all. A scheduling policy must make tradeoffs between these goals.
For example, a scheduling policy that prioritizes interactive tasks to reduce latency may provide
unfair allocations that also lead to starvation. As another example, a scheduler that provides
fair allocation over a small time scale may hurt progress by increasing the number of context
switches. CPU schedulers fall into two broad categories: real-time and best-effort.
Schedulers in the real-time category provide guarantees about how long it will take to respond to
an event; these schedulers ensure the application-defined deadlines are always met. Real-Time
schedulers are typically found in environment requiring latency guarantees, like robotics and
embeded systems. To provide these guatantees, real-time schedulers need to know the CPU
allocation and latency requirements of an application requires, the application is not run. This
admission control policy limits the concurrency of real-time systems.
Best-Effort schedulers, in contrast, provide no guarantees; their primary goal is ease-of-use.
Because they provide only best-effort service, they require no a priori knowledge of application
latency or allocation requirements. Best-effort schedulers also do not have adimission control
mechanisms to prevent CPU contention. These schedulers are found in all commodity operating
systems and used by both desktop and server class machines. Best-effort schedulers are commonly
divided into three groups: time-sharing, proportional-share, and both.