The Dangers and Complexities of SQLite Benchmarking

Benchmarking systems in a repeatable fashion is complex and error-prone. The systems community has repeatedly discussed the complexities of benchmarking and how to properly report benchmarking results. Using the example of SQLite, we examine the current state of benchmarking in industry and academia. We show that changing just one parameter in SQLite can change the performance by 11.8X, and that changing multiple parameters can lead up to a 28X difference in performance. We find that these configuration parameters are often not set or reported in academic research, leading to incomplete and misleading evaluations. Running different off-the-shelf SQLite benchmarking tools such as Mobibench and Androbench in their default configuration shows upto 50% difference in performance. We hope this paper incites discussion in the systems community and among SQLite developers. We hope that our detailed analysis will help application developers to choose optimal SQLite parameters to achieve better performance.