On-line prediction of execution times - a basis for adaptive scheduling

Abstract

Adaptive scheduling meets the need to react in a predictable manner to dynamic, unpredictable changes in the computing system environment. Our TAFT (Time-Aware, Fault-Tolerant) scheduling approach especially addresses the realization of this important property for future real-time applications. The main idea is to trade timeliness of tasks to be executed for maximum functionality. TAFT scheduling uses the concept of the expected case execution time (ECET) of a task in order to achieve adaptivity. We describe the design and the implementation of an integrated on-line monitoring and execution time prediction infrastructure that is able to compute the ECETs efficiently during run-time. We explain how object-orientation can be utilized to achieve more fine-grained, and thus often more precise execution time predictions and we exhibit performance figures, that prove the viability of the approach.