.png){kind=icon}
Return
Hybrid scheduling for serverless function workflows in edge computing system
.png){kind=icon}
PRE
AI
DOI：10.1016/j.future.2026.108822.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Function-as-a-Service (FaaS) computing paradigm enables fine-grained resource scheduling through decomposing complex workflows into a series of functions deployed in edge computing systems by a containerized manner. The temporal constraints of these workflows are required to be satisfied, where delay-sensitive and delay-tolerant workflows are scheduled through a priority-aware mechanism. We focus on resource-constrained edge nodes which are characterized by limited energy supply and computing capability. It is challenging to achieve the optimal trade-off between response delay and energy consumption through optimally deployed functions on edge nodes for scheduling unpredictable incoming workflows. This issue is formulated as an online deadline-aware function deployment model in terms of the Lyapunov optimization technique with efficient stability queues. A Markov decision process is formulated by continuously updating system states, selecting scheduling actions, and evaluating corresponding rewards. We develop a Deep Q-Learning based Workflow Scheduling (denoted as DQLWS) technique to achieve adaptive hybrid workflow scheduling, while considering long-term response delay and energy efficiency. Extensive experiments are conducted in terms of different proportions between delay-sensitive and delay-tolerant workflows, delay-energy weighting settings, scalability of edge node quantity and long-term scheduling performance. Evaluation results show that our approach outperforms the state-of-the-art approaches in multiple comparison dimensions, including response delay, energy consumption, temporal violation rate, and algorithm stability, and thus achieves near-optimal scheduling solutions.
Journal
F
IF:
6.1
Papers:
6.9K
Citations:
2.3W
Organization
View more organizations
Cited Papers
No cited papers available