Modeling Parallel Real Time Workloads on Parallel Systems

A parallel real time system consists of several processors which are used to execute a set of parallel jobs, each of which have an associated deadline. Each job can acquire, use, and release processors dynamically in an effort to complete its execution before its deadline expires. In this paper, a new approach for modeling such systems is presented. In this approach, each job is represented by two independent processes. One represents the service of the job itself, while the other represents the job's deadline expiration. The approach uses Markovian state space models to predict performance of parallel real time systems. The approach presented in this paper provides a uniform framework for evaluating policies for the allocation of processors to jobs. Analytical models are used to compare various traditional policies, including rate monotonic, earliest deadline first, least laxity first, shortest remaining service time first, and equal partitioning. Using the new modeling appr..