Latency-Aware Task Scheduling in Heterogeneous Edge-Cloud Continuum: An Empirical Analysis of Preemptive DAG Partitioning Under Stochastic Workload Volatility

Abstract

Edge-cloud continuum architectures present non-trivial scheduling challenges when directed acyclic graph (DAG)-structured computational workloads exhibit stochastic inter-task dependency volatility. This paper presents an empirical analysis of preemptive DAG partitioning strategies deployed across heterogeneous edge nodes and cloud back-ends, evaluating end-to-end latency, resource utilization efficiency, and fault-tolerance resilience. We instrument a real-world testbed comprising twelve heterogeneous edge devices and a multi-tenant cloud cluster, subjecting it to synthetically generated and trace-driven workloads sampled from production microservice pipelines. Experimental results demonstrate that a latency-aware, criticality-weighted partitioning heuristic reduces makespan by 23.7% and tail latency (P99) by 31.4% over baseline First-Fit and HEFT schedulers, while maintaining energy overhead within acceptable operational margins. These findings establish quantifiable benchmarks for edge-cloud co-scheduling in latency-sensitive industrial IoT deployments.