Skip to Main Content
Resource managers (RMs) often expose configuration parameters that have a significant impact on the performance of the systems they manage. Configuring RMs is challenging because it requires accurate estimates of performance for a large number of configuration settings and many workloads, which scales poorly if configuration assessment requires running performance benchmarks. We propose an approach to evaluating RM configurations called model fuzzing that combines measurement and simple models to provide accurate and scalable configuration evaluation. Based on model fuzzing, we develop a methodology for configuring RMs that considers multiple evaluation criteria (e.g., high throughput, low number of threads). Applying this methodology to the .NET thread pool, we find a configuration that increases throughput by 240% compared with the throughput of a poorly chosen configuration. Using model fuzzing reduces the computational requirements to configure the .NET thread pool from machine-years to machine-hours.