SLA aware deep reinforcement learning for adaptive EdgeCloud task scheduling

The highly dynamic and hybrid workloads executed in edge–cloud systems lead to many deadline misses, higher latency, and energy waste, as a single static scheduler is usually unable to meet Service Level Agreements (SLAs). To overcome the practical gap, a framework is proposed by employing SLA-aware deep reinforcement learning (SLA-DRL) to schedule tasks under the SLAs as a first-class objective while still optimising for latency and energy. The proposed approach categorizes each incoming task into an appropriate Service Level Agreement (SLA) tier—Gold, Silver, or Bronze—and computes a task-level SLA Violation Risk Score (SVRS). This score is derived from the task deadline, prevailing system load conditions, and recent SLA violation history. Further, it assists the agent concentrate on tasks with a higher probability of skipping SLA. The proposed lightweight SLA-aware action pruning module is capable to prune placement actions that are clearly unsafe and also it not requires evaluation by the policy network, thereby stabilising learning and enhances convergence speed. To validate it, a trace-driven simulation of a heterogeneous edge–cloud cluster is designed and evaluated the proposed SLA-DRL against existing schedulers (FIFO, Round Robin, Earliest Deadline First) as well as a non-SLA DRL version. The results indicate that proposed approach reduces SLA violations by 41.6%, average latency by 32.1%, and energy consumption by 28.5% compared to the best baseline, with similar gains across all experiments with different workload mixes and priority levels. This shows that explicitly using SLA information in the DRL state, reward, and action space can create a more robust and adaptive scheduler that better matches upcoming edge–cloud applications.