We have seen above several different schemes for processor scheduling. These can be distinguished by how they handle three orthogonal issues:
Do they support space scheduling or time scheduling?
Do they support dynamic or static partitions, that is, do they keep the number of schedulable virtual processors of an application static or dynamic? A policy is more dynamic than another if it changes these partitions more frequently.
Do they coordinate or not with the applications when they reallocate processors, that is, use information provided by the application (e.g. whether a VP is at a safe point, whether a thread is executing a critical region) when they reallocate?
The following table describes how the four policies we have seen handle these issues:
Equipartition and Process Control are not distinguished by this taxonomy because it does not distinguish between different coordination schemes. They are Quasi-Static since even though an application's need (number of VPs it creates) is static, its quota (number of unsuspended VPs) is dynamic and changes as applications are created/killed. As a result, the partition of an application changes less frequently under McCann's Dynamic policy.
McCann et al evaluated these policies based on the following criteria:
Response Time
Fairness
Response Time to Short Jobs
In all of their experiements, they used 16 processor machines.
Let us consider each in turn.