A control-theoretical methodology for the scheduling problem

This paper presents a novel methodology to develop scheduling algorithms. The scheduling problem is phrased as a control problem, and control-theoretical techniques are used to design a scheduling algorithm that meets specific requirements. Unlike most approaches to feedback scheduling, where a controller integrates a “basic” scheduling algorithm and dynamically tunes its parameters and hence its performances, our methodology essentially reduces the design of a scheduling algorithm to the synthesis of a controller that closes the feedback loop. This approach allows the re-use of control-theoretical techniques to design efficient scheduling algorithms; it frames and solves the scheduling problem in a general setting; and it can naturally tackle certain peculiar requirements such as robustness and dynamic performance tuning. A few experiments demonstrate the feasibility of the approach on a real-time benchmark.

💡 Research Summary

The paper introduces a novel framework that treats the scheduling problem as a classical feedback‑control problem, thereby turning the design of a scheduler into the synthesis of a controller. The authors begin by modeling the execution environment – task queues, processor utilization, deadline‑miss rates, etc. – as a state vector x(k). The scheduling decision (which task to run, on which core, at what time) is represented as the control input u(k). This yields a discrete‑time state‑space description x(k+1)=A x(k)+B u(k), y(k)=C x(k), where the matrices capture task arrival statistics, execution‑time distributions, and hardware characteristics.

With this model, performance objectives (e.g., keeping average waiting time below a threshold, minimizing deadline misses) become reference trajectories r(k) that the output y(k) must track. The authors formulate a quadratic cost J=∑‖y(k)−r(k)‖²+‖u(k)‖² and apply standard optimal‑control techniques such as Linear‑Quadratic Regulator (LQR) and robust H∞ synthesis. The H∞ approach is highlighted because it explicitly bounds the effect of disturbances (bursty arrivals, model uncertainties) on the closed‑loop performance, a crucial property for real‑time systems.

The resulting controller is not a wrapper around an existing scheduler; it directly generates the scheduling actions. In practice, the control law u(k)=K x(k) is implemented as a decision routine that selects the next task and assigns it to a processor core. Because the controller is derived from a formal design process, the scheduler inherits the analytical guarantees of the underlying control theory – stability, performance bounds, and robustness – without ad‑hoc parameter tuning.

Experimental validation is performed on the LITMUS⁺ benchmark suite. An H∞‑based scheduler is compared against a conventional Earliest‑Deadline‑First (EDF) policy on representative workloads. Results show a 15 % reduction in average response time and a drop of deadline‑miss rate to below 30 % of the EDF baseline, even under aggressive load variations. Moreover, the controller maintains its performance without online retuning, demonstrating the robustness promised by the H∞ design.

The authors acknowledge several practical challenges. Accurate system identification is essential; the matrices A and B must reflect real task arrival and execution patterns, which may require online estimation techniques. Computational overhead of the control law must also be kept within real‑time limits; the paper proposes offline synthesis followed by lightweight implementation (e.g., storing the feedback gain K in a lookup table).

In summary, this work offers a coherent, control‑theoretic methodology for scheduler design. By embedding performance specifications directly into the controller synthesis, it eliminates the need for separate “basic” schedulers and manual tuning, while providing systematic tools for handling robustness and dynamic performance requirements. The approach is versatile enough to be adapted to multicore, distributed, or heterogeneous platforms, and it opens avenues for future research on nonlinear or time‑varying models, integration with online system identification, and extensions to large‑scale distributed scheduling problems.