LLM INFERENCE PERFORMANCE ANALYSIS ACROSS WINDOWS, WINDOWS SUBSYSTEM FOR LINUX 2 AND NATIVE LINUX

Abstract

The increasing use of Large Language Models (LLMs) on consumer devices raises practical concerns regarding inference latency and resource efficiency, which can vary depending on the operating system environment. However, it is still unknown which operating system is most optimal and efficient for executing local LLM workloads on hardware with limited resources. Therefore, this study aims to compare and evaluate the inference performance of LLMs across three operating system environments: native Windows, Windows Subsystem for Linux 2 (WSL2), and native Linux on a low-spec laptop equipped with a 12th Gen Intel Core i3 processor and 8 GB of RAM. The Ollama framework is used to run several lightweight LLM models, including Llama3.2:3B, Qwen2.5:3B, and Phi3.5:3.8B. An experimental benchmarking approach is conducted using three levels of prompt complexity (light, medium, and heavy), with 10 trials for each prompt scenario to ensure consistency of results. Evaluation metrics include inference latency, CPU utilization, and memory consumption during model execution. Results show that the operating system environment significantly impacts the inference performance of LLMs. For example, using the Llama3.2:3B model under heavy prompting, the average latency reaches 71.24 seconds on Windows, 75.96 seconds on WSL2, and 70.80 seconds on native Linux, while under light prompting the average latency is 11.76 seconds, 12.73 seconds, and 10.83 seconds, respectively. These findings indicate that native Linux generally provides more efficient LLM inference on resource-constrained hardware. This study contributes an empirical and repeatable comparison across operating system environments, providing practical insights for selecting deployment platforms for LLM applications.