next up previous
Next: Events Up: Kernel Support Previous: Virtual-Processor Model

Domain Scheduling

To explain the scheduling mechanism adopted in Nemesis requires an understanding of how we see a flexible multimedia platform being used. The allocation of resources to applications will not be controlled soley by the applications themselves. Rather we see users being able to control processor allocation much in the same way that they control pixel allocation in window systems. Thus applications will not always get what they want; they will have to adapt to the resources they are given. However, for a particular time, seconds or tens of seconds, some of the resources given to an application may be viewed as ``guaranteed''. The application may choose to use an particular algorithm on the basis of this guarantee. It may also be able to exploit unguaranteed resources which become available fortuitously. The approach to scheduling in Nemesis is to schedule domains with a weighted scheduling discipline, where the weights are calculated from the user's current policy. Within a given time frame, not all domains may use their allocation; the policy for sharing out remaining resources is still the subject of investigation. While domains have some processor allocation remaining, the current scheduler implementation uses an earliest deadline first algorithm to select between them.

Above this primitive-level scheduler, and running on a longer time scale is a Quality-of-Service-manager domain whose task is to update the scheduler weights; this is performed not only in response to applications entering or leaving the system, but also adaptively as applications modify their behaviour -- this is performed on a longer time scale that the individual scheduling decisions in order to smooth out short-term variations in load.

Sape J. Mullender, Ian M. Leslie and Derek McAuley