SETS: Stochastic execution time scheduling for multicore systems by joint state space and Monte Carlo

The advent of multicore platforms has renewed the interest in scheduling techniques for real-time systems. Historically, `scheduling decisions' are implemented considering fixed task execution times, as for the case of Worst Case Execution Time (WCET). The limitations of scheduling considering WCET manifest in terms of under-utilization of resources for large application classes. In the realm of multicore systems, the notion of WCET is hardly meaningful due to the large set of factors influencing it. Within soft real-time systems, a more realistic modeling approach would be to consider tasks featuring varying execution times (i.e. stochastic). This paper addresses the problem of stochastic task execution time scheduling that is agnostic to statistical properties of the execution time. Our proposed method is orthogonal to any number of linear acyclic task graphs and their underlying architecture. The joint estimation of execution time and the associated parameters, relying on the interdependence of parallel tasks, help build a `nonlinear Non-Gaussian state space' model. To obtain nearly Bayesian estimates, irrespective of the execution time characteristics, a recursive solution of the state space model is found by means of the Monte Carlo method. The recursive solution reduces the computational and memory overhead and adapts statistical properties of execution times at run time. Finally, the variable laxity EDF scheduler schedules the tasks considering the predicted execution times. We show that variable execution time scheduling improves the utilization of resources and ensures the quality of service. Our proposed new solution does not require any a priori knowledge of any kind and eliminates the fundamental constraints associated with the estimation of execution times. Results clearly show the advantage of the proposed method as it achieves 76% better task utilization, 68% more task scheduling and deadline miss reduction by 53% compared to current state-of-the-art methods.