The Tail Latency Paradox: CPU Affinity and Scheduling under Host CPU Saturation

As containerized edge workloads increasingly deploy on desktop-class host operating systems and hypervisors, understanding the interactions between operating system scheduling policies and microservices network Quality of Service (QoS) becomes critical. This paper evaluates the effectiveness of CPU core pinning (taskset) and real-time process scheduling (SCHED_FIFO) in suppressing network tail latency under host CPU saturation within a WSL2 container environment. Across N_runs = 5 randomized, independent trials (180s each), baseline Completely Fair Scheduler (CFS) execution achieved an average round-trip time (RTT) of 0.793 ± 0.007 ms under 100% CPU stress, with P_95 and P_99 tail latency bounds of 2.000 ± 0.000 ms and 2.042 ± 0.094 ms, respectively. Surprisingly, applying CPU core isolation (Iso_CFS) or real-time priority scheduling (Hybrid_Iso_RT) failed to compress network tail latency, extending P_99 bounds to 2.400 ± 0.548 ms and 2.442 ± 0.6101 ms. While telemetry confirms that real-time scheduling completely eliminates user-space involuntary context switches (Δ = 0.000), this process-level preemption isolation does not translate to lower network tail latency. Non-parametric effect size analysis (|δ| ≤ 0.051) confirms a fundamental cross-layer disconnect: preventing user-space preemption cannot compensate for lower-level kernel softirq queueing delays and hypervisor traversal overheads under host saturation.