End-to-end performance control for distributed real-time systems

A best effort scheduling algorithm is used to reduce the probability of exceeding a deadline-related target time for each job class allocated to a distributed real-time system. A dynamic scheduler is used as a point of control on each node of a pipeline-structured distributed system. The authors contrast a local clustering scheduling algorithm which attempts to meet end-to-end timing requirements in isolation to a global clustering scheduling algorithm which attempts to meet end-to-end timing requirements by utilizing information about the state of other nodes in a job´s execution path. The approach incorporates both job-class-based loss functions and feedback of remaining time estimates. A simulation study has shown that the global algorithm can provide an even distribution of processing power over all nodes and job classes while maintaining system stability