Learning to Optimize Workflow Scheduling for an Edge–Cloud Computing Environment

Abstract

The widespread deployment of intelligent Internet of Things (IoT) devices brings tighter latency demands on complex workload patterns such as workflows. In such applications, tremendous dataflows are generated and processed in accordance with specific service chains. Edge computing has proven its feasibility in reducing the traffic in the core network and relieving cloud datacenters of fragmented computational demands. However, the efficient scheduling of workflows in hybrid edge–cloud networks is still challenging for the intelligent IoT paradigm. Existing works make dispatching decisions prior to real execution, making it difficult to cope with the dynamicity of the environment. Consequently, the schedulers are affected both by the scheduling strategy and by the mutual impact of dynamic workloads. We design an intelligent workflow scheduler for use in an edge–cloud network where workloads are generated with continuous steady arrivals. We develop new graph neural network (GNN)-based representations for task embedding and we design a proximal policy optimization (PPO)-based online learning scheduler. We further introduce an intrinsic reward to obtain an instantaneous evaluation of the dispatching decision and correct the scheduling policy on-the-fly. Numerical results validate the feasibility of our proposal as it outperforms existing works with an improved quality of service (QoS) level.