Windows Process Priority and Local LLM Inference: A Counterbalanced Host-Runtime Study

Abstract

This technical note quantifies the practical effect of the standard Windows process-priority mechanism on local large-language-model inference throughput. The experiment was conducted with Local Frontier Orchestrator (LFO) v9.6.3 on a constrained Windows 10 CPU-only host using Ollama and a Qwen 3.5 4B model profile. An exploratory pilot was followed by a six-block counterbalanced comparison of Windows Normal and High process priority. The confirmation experiment comprised 36 measured LOCAL-only samples and 18 matched block-by-prompt pairs. Mean prefill throughput increased from 18.7034 to 22.0197 tokens/s (+17.7%), with High priority outperforming Normal in all 18 matched pairs. Mean decode throughput increased from 3.8150 to 4.1036 tokens/s (+7.6%), but this effect was smaller and noisier, with High faster in 12 of 18 pairs. The contribution is not the observation that Windows High priority receives preferential CPU scheduling, which is documented operating-system behavior, but the empirical measurement of how strongly that standard scheduling mechanism translated into local-LLM prefill and decode throughput under a reproducible workload. The result supports treating host-runtime scheduling conditions as part of local-LLM benchmarking and runtime configuration, particularly for CPU-involved inference workloads. The evidence is intentionally host-scoped. It does not establish a universal Windows speedup, does not identify a specific microarchitectural mechanism, and does not claim replication of the Raspberry Pi memory-wall mechanism that motivated the broader host-runtime question. The Zenodo record includes the main technical note together with machine-readable research data and supplementary reproducibility materials. The archived dataset contains the pilot and counterbalanced measurements, matched-pair tables, process-priority verification, runtime placement records, provenance information, and checksums. Supplementary materials include publication figures, figure-generation code, the frozen statistical analysis contract, canonical statistics, claim-boundary documentation, and reproduction scripts. Related work includes Daniel Correa Villa, “Pushing Four Raspberry Pis to the Memory Wall: Bit-Exact 30B Mixture-of-Experts Inference, +16% over the Public Record” (DOI: 10.5281/zenodo.20357376), and the earlier LFO transferability technical note “Reducing Context and Generation Traffic in a Local-First LLM Orchestrator: A Transferability Case Study” (DOI: 10.5281/zenodo.23318551).