Dynamic Resource Scheduling for Real-Time Multimodal Inference on Heterogeneous Edge Devices: A Trace-Calibrated Study of Deadline-Aware Audio–Visual Scheduling

Real-time multimodal inference couples resource-intensive visual and acoustic feature extraction with latency-sensitive fusion. Heterogeneous edge nodes can offer unequal execution times, while static routing ignores queue states and request deadlines. This paper investigates a low-overhead dynamic routing scheme, EDF-ECT, which combines estimated completion-time placement with non-preemptive earliest-deadline-first local queues. A compact dual-branch convolutional graph is timed on a CPU using three threading configurations; 432 measured invocations calibrate three virtual service profiles. An event-driven replay evaluates six policies using workload-adjusted arrival rates and common per-request service samples. Across eight independent traces per setting at nominal load 0.95, EDF-ECT improves the deadline-hit rate from 94.57% to 97.07% for balanced traffic and from 96.92% to 99.27% for large-input-heavy traffic. Paired 95% confidence intervals for the gains are [1.73, 3.26] and [1.96, 2.75] percentage points. Deadline gains trade off against unconditional p95 response time. The findings provide a reproducible request-level scheduler study with measured service kernels while distinguishing virtual workers from physical edge devices.