Implementing real-time systems using performance polymorphism

A novel model for complex real-time systems is proposed. In this model, several versions of a program fragment are provided to perform a particular action. These versions will differ only in their performance parameters, such as the time required, the resources consumed, and the precision of the results. The authors describe an implementation of a technique called performance polymorphism, in which the process of selecting a version from this set may be automated. Performance polymorphism is a unified theory to express the choice among multiple versions in a way that is both natural and powerful. It allows the flexibility of adding new versions at any time, of adapting to unforeseen constraints, and of adapting to automatically generated variants of a procedure (as, for example, might come from a parallelizing compiler). A means to implement the theory of performance polymorphism that requires very low overheads at run time has been developed by analyzing the constraints and propagating performance information backwards in order to bind performance polymorphic functions as early as possible.