In this research, a new QoS-aware Dynamic Scheduling and Resource Allocation (QoS-DASRA) framework has been suggested to be implemented in intelligent ML inference management for the edge, fog, and cloud levels. In this approach, the first step involves developing a QoS Satisfaction Model (QSM), which includes latency, deadline, computing needs, bandwidth, reliability, and energy parameters. Subsequently, the dynamic task placement scheme considers the appropriateness of the candidate resource based on the QoS-resource performance metric in real-time. Finally, an adaptive resource allocation algorithm optimizes CPU, memory, bandwidth, and computing capability based on the workload. At 30 concurrent inference tasks, QoS-DASRA is achieving 96 ms average latency, 92.1% QoS satisfaction, 88% resource utilization, 7.9% deadline violation rate, and 125 J energy consumption. when it is compared with the strongest baseline, the architecture is reducing the latency by 26.2%, is increasing QoS satisfaction by 8.3%, is reducing the deadline violations by 51.2%, and is decreasing energy consumption by 15.5%. The results are showing that QoS-DASRA is providing adaptive and efficient resource management for ML inference under the varying edge-fog-cloud workload conditions.

Computing, Fog Computing, QoS-Aware Scheduling, Machine Learning Inference, Dynamic Resource Allocation