Dynamic Scheduling of Aperiodic Real-time Task Graphs on Heterogeneous Edge Computing Environments

Abstract

The support for Multi-access Edge Computing (MEC) along with Quality of Service (QoS)-class specific communications in 5G+ mobile networks, are opening up horizons for new applications and business models. For example, this has ushered in the possibility of remotely executing real-time automated monitoring and control applications as edge services. Such services can be offered by futuristic carrier-network operated MECs with latency-sensitive communication capability. Due to the complex interactions among their functionalities, these real-time control applications are often modeled as Precedence-constrained Task Graphs (PTGs). Although, a majority of these applications are time-triggered and hence periodic in nature, many of them may be dynamically spawned being triggered by events such as alarms. In such PTGs, tasks may have multiple quality levels, with higher levels offering greater accuracy or reliability, enhancing application QoS. This paper proposes a QoS-aware any-time methodology for dynamically scheduling an aperiodic PTG-structured application on an MEC system that is partially preoccupied by persistent, periodic real-time workloads. A two-phased heuristic list-scheduling strategy called Dynamic QoS Aware DAG Scheduler (D-QADS) has been proposed for this purpose. In the first phase, D-QADS generates a base schedule using a strategy derived by extending the well-known Heterogeneous Earliest Finish Time ( HEFT ) list scheduler, to enable schedule generation on partially pre-occupied systems. Subsequently, in the second phase, this base schedule is improved by iteratively enhancing the QoS levels of tasks. Extensive simulation-based experiments with both randomly generated as well as standard benchmark DAGs, along with comparisons to a baseline approach, Greedy D-QADS ( GD-QADS ), an extended version of the state-of-the-art HEFT algorithm, exhibited the efficacy of our algorithm in diverse scenarios. Results show that D-QADS can achieve up to \(14 \% \) more QoS than GD-QADS . The algorithm is fast, reaching the maximum achievable QoS within 8.4 ms for 20-task PTGs on a 30% preloaded four-processor system.