Architectural support for greater predictability in real-time systems

Abstract

A real-time system is characterized by computational tasks that need to complete within well-defined time durations. Scheduling policies help guarantee deadlines, but scheduling policies and the analysis associated with them need exact information to guarantee timing correctness. This exact information includes the worst-case execution time of tasks in the system. Existing architecture techniques (caches, superscalar out-of-order execution) improve the average case performance of a computer system but complicate the estimation of worst-case execution times. In the talk I will discuss some recent techniques for estimating WCETs and some techniques for more predictable execution, including the use of scratchpad memories.