Tuning a software system’s configuration is essential to meet performance requirements. However, not only do configuration options affect performance, but also the system’s interaction with external factors such as the workload. Hence, tuning requires understanding how a specific setting of external factors (e.g., a specific workload) in combination with the system configuration influences performance. Current performance modeling approaches usually do not incorporate external factors, for good reasons: training a separate model per setting is costly and is unlikely to generalize, whereas a single model trained on multiple settings fails to capture variations that are specific to a certain setting.

To address this shortcoming, we propose HyPerf, a Bayesian multi-level performance modeling approach that systematically distinguishes between setting-invariant and setting-variant influences, that is, influences that remain consistent across settings versus those that exhibit substantial variation. For this purpose, HyPerf employs a hierarchical structure: The upper level captures general performance trends across multiple settings (e.g., across different workloads), while the lower level refines these estimates with setting-specific deviations (e.g., workload-specific performance variations).

With HyPerf, we aim at balancing accuracy and efficiency, achieving robust performance predictions with significantly fewer training samples. Unlike the state of the art, HyPerf is able to identify a minimal set of settings that captures essential performance variations, so that developers can approximate whether all setting-variant influences have been accounted for.

Empirical evaluations on ten real-world software systems across up to 35 workloads demonstrates that HyPerf matches or outperforms state-of-the-art approaches while requiring fewer measurements. Notably, HyPerf enables interpretable performance reasoning and can identify minimal workload subsets that capture essential performance variations.