Improving rewards in overloaded real-time systems

TL;DR

This paper introduces a workload-informed scheduling policy for overloaded real-time systems that improves revenue, supported by theoretical proof and empirical evidence showing its superiority over static and other dynamic policies.

Contribution

A novel stochastic approximation-based scheduling policy that leverages workload distribution to enhance revenue in overloaded real-time systems.

Findings

The proposed policy increases revenue compared to static resource allocation.

Empirical results show the policy outperforms other scheduling strategies.

The approach is applicable to various soft real-time systems with multiple service classes.

Abstract

Competitive analysis of online algorithms has commonly been applied to understand the behaviour of real-time systems during overload conditions. While competitive analysis provides insight into the behaviour of certain algorithms, it is hard to make inferences about the performance of those algorithms in practice. Other approaches to dealing with overload resort to heuristics that seem to perform well but are hard to prove as being good. Further, most work on handling overload in real-time systems does not consider using information regarding the distribution of arrival rates of jobs and execution times to make scheduling decisions. We present an scheduling policy (obtained through stochastic approximation, and using information about the workload) to handle overload in real-time systems and improve the revenue earned when each successful job completion results in revenue accrual. We prove that the policy we outline does lead to increased revenue when compared to a class of scheduling policies that make static resource allocations to different service classes. We also use empirical evidence to underscore the fact that this policy performs better than a variety of other scheduling policies. The ideas presented can be applied to several soft real-time systems, specifically systems with multiple service classes.