Deadline-aware scheduling mechanisms become in-effective under severe hardware interference during overload conditions. In edge environments where AI inference and coordi-nation workloads are co-located on the same node, contention in the shared Last-Level Cache (LLC) substantially increases effec-tive service times irrespective of the underlying scheduling policy, resulting in uncontrolled queue buildup and catastrophic tail-latency degradation that cannot be mitigated through priority-based ordering alone.

We introduce MOSAIC, a PMU-guided overload isolation framework designed for mission-critical edge computing sys-tems. MOSAIC leverages hardware Performance Monitoring Unit (PMU) counters to quantify pairwise LLC interference across workload classes and incorporates these measurements into a proactive admission-control mechanism. In addition, a predictive deadline-feasibility metric guides scheduling decisions while enforcing bounded starvation guarantees. Collectively, these mechanisms shield mission-critical workloads from harmful co-location effects during catastrophic burst conditions.

Experimental results demonstrate that MOSAIC reduces P99 tail latency by 26.4% compared with the strongest priority-aware baseline, while simultaneously achieving 100% completion for mission-critical task classes under crisis-level workloads. The framework incurs minimal runtime overhead, with PMU sampling requiring only 1.4 µs per read and admission-control decisions completing within 280 µs. Furthermore, the system is fully reproducible and publicly released as open-source software.