Multi-Intention-Aware Configuration Selection for Performance Tuning

Configuration tuning can improve software performance. Pre-selecting performance-related parameters can significantly reduce the search space during tuning. These works, however, are both limited by the specific workloads chosen to train their models. More importantly, they are unaware of user intentions other than performance but are also important (e.g., reliability, security). Given these limitations, we find that the configuration document often (even if it does not always), contains rich information about the parameters' relationship with many user intentions. However, documents might also be long and domain specific. Thus, we focus on guiding users in selecting performance-related parameters while warning about side-effects on non-performance intentions via mining documents. In this paper, we first conduct a comprehensive study on 13 representative software containing 7,325 configuration parameters, and derive six types of ways in which configuration parameters may affect non-performance intentions. Guided by this study, we design SafeTune, a workload-independent method that pre-selects important performance-related parameters and warns about their sideeffects on non-performance intentions. Evaluation on target software shows that SafeTune correctly identifies 6-22 performancerelated parameters that are missed by state-of-the-art tools but have significant performance impacts (up to 14.7x). Furthermore, our case study demonstrates that SafeTune can successfully help the