Cache memory management in real-time systems

Abstract

Dependable real-time systems are essential to time-critical applications. The systems that run these applications require high degrees of performance and predictability. Although memory caching has long been known as a means of increasing system performance, it is typically unpredictable in nature.

This thesis presents a technique called preferred preemption points to increase the predictability and performance of cache-based real-time systems. A performance analysis tool and methodology designed to select and evaluate preemption points are presented with the technique. The technique increases the ability to calculate system preemption overhead, leading to greater determinism and consequently performance, and providing for tighter bounds on calculated worst-case execution times. The performance of preferred preemption points is demonstrated using the analysis tool. Results are provided for various system configurations. System configurations are defined by cache characteristics and preemption frequencies. In some instances results have shown a 10% reduction in execution time with preferred preemption points. Such savings in execution time provide greater flexibility in scheduling and increase throughput of time-critical tasks. The thesis addresses the issue of including preferred preemption points into real-time scheduling algorithms. Other system considerations with regard to preferred preemption points are also considered.