A Transformer-based Approach to Real-Time Task Scheduling in Latency-Critical Edge Systems
PDF

Keywords

Edge computing
Task scheduling
Transformer architecture
Real-time systems
Latency optimization
Self-attention mechanism

Abstract

Edge computing has emerged as a critical paradigm for supporting latency-sensitive applications by
bringing computational resources closer to data sources. However, efficient task scheduling in edge
systems remains a significant challenge due to the heterogeneous nature of resources, dynamic
workload patterns, and stringent latency requirements. Traditional heuristic-based scheduling
algorithms often struggle to adapt to the rapidly changing conditions inherent in edge environments.
This paper presents a novel Transformer-based approach for real-time task scheduling in latencycritical edge systems. By leveraging the self-attention mechanism of Transformer architecture, our
proposed method learns complex temporal dependencies and spatial relationships among tasks and
resources, enabling intelligent scheduling decisions that minimize response latency while
maintaining high resource utilization. The methodology incorporates a multi-head attention
mechanism to capture diverse task characteristics including priority levels and deadline constraints,
positional encoding to preserve temporal information, and a reinforcement learning framework to
optimize long-term performance. Experimental results demonstrate that the proposed approach
achieves up to 37% reduction in average task completion time and 42% improvement in deadline
satisfaction rate compared to state-of-the-art baseline algorithms including completion time-based
assignment with hierarchical weighted assignment and weighted round-robin scheduling methods.
The system maintains deadline satisfaction rates above 94% for time-critical tasks even under high
workload intensities, adapting effectively to resource heterogeneity across edge-cloud continuum
architectures. 

PDF
Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 International License.

Copyright (c) 2025 Klara Vuković, Alessandro Romano (Author)