|
UW–MADISON · JAN.–MAY 2026 How do background file reads affect an otherwise fixed inference workload? Honors course research advised by Prof. Michael Swift.  I evaluated a fixed LLM inference workload under controlled background filesystem reads. Median inference latency increased 17.1% relative to the baseline. Windows PerfMon showed stable working sets and page-fault counts plateauing near 29.5K even as latency rose. These observations support shared-resource contention over working-set thrashing as the explanation; they do not isolate a single hardware bottleneck. The study connects operating-system activity to inference performance: controlling the model workload alone is insufficient when it shares resources with other processes. |