A performance model for parallel real-time systems

Real-time systems need to exploit high-speed parallel processing technology to meet their strict timing requirements. A general parallel architecture may have multiple stages connected in a pipeline, where each stage has multiple servers. This thesis deals with performance analysis of such parallel real-time systems. A regular pipeline model is proposed as a mathematically tractable model. A concept of optimal instance blocking is introduced to achieve flow balance at stage. A minimum-response-priority-assignment algorithm is presented to achieve the minimum end-to-end response time for each task. A concept of instance distribution is introduced to schedule a task whose execution time is greater than period. A general model for a pipelined multi-server parallel real-time system is developed in stages, starting from a single server model. Each modelling stage uses results from the previous stages. The final performance model provides a schedulability test; calculates end-to-end task response times, required buffer and optimal hardware capacity; and includes an optimal scheduling algorithm for each stage. The model has been validated by simulation. The limitations of the model are discussed. The model is theoretical rather than practical to achieve mathematical tractability. However it gives us predictable upper-bound estimates for the system performance.