The Tail Latency Paradox: Why OS Scheduling Isolation Fails Under Host CPU Saturation

Official open-source repository containing replication scripts, telemetry suites, and raw experimental data for the paper:

Yiu, T. W. (2026). The Tail Latency Paradox: Why OS Scheduling Isolation Fails Under Host CPU Saturation. Zenodo. https://doi.org/10.5281/zenodo.21896255

---

Abstract & Key Findings

As containerized microservices share physical host hardware, managing network Quality of Service (QoS) tail latency is critical. This empirical evaluation investigates whether OS-level preemption isolation—specifically CPU core pinning (taskset) and real-time priority scheduling (SCHED_FIFO)—can mitigate network latency spikes during 100% CPU saturation within container environments.

Core Insight

While real-time scheduling completely eliminates user-space involuntary context switches ($\Delta = 0.000$), it fails to compress {99}$ network tail latency. Non-parametric effect size analysis ($|\delta| \le 0.051$) reveals a fundamental cross-layer disconnect: user-space preemption isolation cannot bypass lower-level kernel softirq packet processing bottlenecks and hypervisor queueing delays.

---

Benchmark Results ({\text{runs}} = 5$, 100% CPU Stress)

Configuration	Isolation Method	Mean RTT (ms)	{95}$ Latency (ms)	{99}$ Latency (ms)
Shared CFS (Baseline)	Completely Fair Scheduler	-bash.793 \pm 0.007$	.000 \pm 0.000$	.042 \pm 0.094$
Iso CFS (Core Pin)	Core Pinning (taskset)	-bash.930 \pm 0.446$	.400 \pm 0.548$	.400 \pm 0.548$
Shared RT (SCHED_FIFO)	Real-Time Priority	-bash.852 \pm 0.255$	.600 \pm 0.548$	.200 \pm 0.447$
Hybrid Iso RT (Combined)	Core Pinning + SCHED_FIFO	-bash.814 \pm 0.378$	.200 \pm 0.447$	.442 \pm 0.610$

---

Repository Structure

.
├── scripts/       # Docker execution, CPU pinning, and workload generation scripts
├── data/          # Raw CSV telemetry logs across all experimental runs
├── paper/         # Manuscript PDF, LaTeX sources, and telemetry plots
└── README.md      # Project overview and reproduction guidelines

---

Citation

If you use this benchmark suite or build upon these findings, please cite the paper using the following BibTeX entry:

@article{yiu2026taillatency,
  author    = {Yiu, Tsun Wa},
  title     = {The Tail Latency Paradox: Why OS Scheduling Isolation Fails Under Host CPU Saturation},
  year      = {2026},
  journal   = {Zenodo},
  doi       = {10.5281/zenodo.21896255},
  url       = {https://doi.org/10.5281/zenodo.21896255}
}

License

Distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) License.