Performance analysis for design of distributed real-time systems with preemptive priority scheduling
D.R. Lefebvre · 1996
Unlike a conventional distributed system that primarily manages information, a distributed real-time system interacts with its environment through sensors and actuators, and typically performs a control function on a physical system. These systems are inherently complex and designing them is a difficult task; also, the technology is relatively new and few standards or design tools have been developed. The thesis presents a new performance analysis approach, named the DIRECS (DIstributed REal-time Constraint Satisfaction) method, for designing distributed systems with hard real-time response requirements. In broad terms, the DIRECS method is an extension of Generalized Rate Monotonic Analysis to distributed systems, although system specification and treatment of event-triggered tasks is distinctly different from that of CMU's Software Engineering Institute. The common ground is the dependence on characteristics of preemptive priority-based scheduling to predict worst case response times for task executions and message transmissions. Features unique to the new method include: analysis of the real-time performance of message queues, inclusion of a large fraction of event-triggered tasks, and modeling of multiple concurrent system responses. Compared to alternative performance analysis methods of simulation, stochastic modeling and semantic modeling, the scheduling-based DIRECS method is more computationally efficient, provides better prediction of hard real-time responses, and is tractable for large systems, respectively. Upper bounds on system response time are predicted by solving a set of constraint equations that define relations between hardware capabilities, software computing requirements, and system response times. The system of constraint equations is under-specified; thus, by adding a cost function the solution can be directed toward different objectives to minimize response times or minimize hardware cost. The problem is solved through a custom successive linear programming algorithm. Experimental verification on a mobile robot control application has demonstrated that the DIRECS method predicts tight upper bounds on system response times in good agreement with measured maximum times and has dramatically lower computational requirements than simulation.