Real-Time Scheduling Optimization via Physical Core Isolation and Hierarchical Task Classification on ARM-Linux

With the increasing demand for hard real-time performance in embedded systems (such as autonomous driving, industrial control), the problem of mutual interference caused by real-time and non-real-time tasks sharing CPU cores in the standard Linux kernel has become increasingly prominent, leading to uncontrollable increases in scheduling latency. To address this, this paper innovatively proposes and validates a hierarchical CPU resource partitioning mechanism for Linux systems on ARM multi-core platforms. This mechanism divides CPU cores into Real-Time cores (RT-CPU) and Non-Real-Time cores (NRT-CPU). By classifying tasks into three levels (hard real-time rt0, soft real-time rt1, non-real-time nrt) and deploying them onto dedicated cores (physical isolation), combined with a dynamic priority-sensitive migration strategy (e.g., migrating rt1 tasks to a high-priority queue on NRT-CPU when RT-CPU load >80%, forcibly migrating nrt tasks out when encountering real-time tasks on RT-CPU), it effectively reduces task execution latency and enhances deterministic guarantees for hard real-time tasks. Comparative experiments conducted on a Raspberry Pi 4B platform using cyclictest demonstrate that: compared to the standard kernel (Linux 5.10_stand), the optimized kernel (cs Linux 5.10_rt) reduces the minimum latency by 44%, decreases the average latency by 2%, and significantly compresses the maximum latency jitter; compared to the PREEMPT-RT patched kernel, it improves the minimum latency by 36%. This mechanism successfully eliminates direct interference from non-real-time tasks, providing reliable support for high real-time scenarios.