Fuzzy scheduling of real-time ensemble systems

This paper addresses the problem of resource scheduling in real-time ensemble systems. An ensemble system uses multiple simple computational models (called “experts”) to produce its outputs. Real system requirements of ensemble systems (e.g., size, weight, power and cost constraints) often lead to limited availability of computational resources required to support concurrent execution of all their experts. In practical systems, uncertainties in execution time and resource utilization complicate even further the scheduling of these experts. We propose a fuzzy-logic feedback-based resource scheduler (FuzzyFES) that can provide real-time execution of all relevant experts while minimizing the impact of limited resources and uncertainties on the system performance. FuzzyFES consists of a fuzzy-logic controller (FZ), a task utilization adaptor (TUA) and a real-time task scheduler (RTS) working harmoniously in a closed loop with an ensemble system to be scheduled. By considering the uncertainties that may be present in the systems and deployment environments, FZ determines the total allowable CPU utilization for the ensemble system. TUA then calculates the amount of resource utilization to be allocated to each expert not exceeding the total allowable utilization. The assigned utilization from TUA ensures that critical experts achieve their best performance while guaranteeing minimum execution time needed by others. RTS creates a real-time schedule for the experts to execute on multiple processors according to the allotted utilization. Our performance evaluation of a case-study ensemble system with limited resources demonstrates that FuzzyFES can schedule experts to produce outputs closely similar to those of the same system with sufficient resources, although the limited-resource system has up to 40% fewer resources. The results also confirm FuzzyFES's efficiency and show that execution-time imprecision and occasional fluctuation of resource availability can be tolerated by at least 45% more than when the experts are scheduled in an open-loop manner.