Using Autoencoders in homogeneous and heterogeneous transfer learning configuration auto-tuning problems

Performance auto-tuning is a trending topic in ICT applications. Historically the optimization of complex systems has been performed by an human actor, but with the growing complexity of current day information systems and the need of fast and effective optimization there is a strong need to automate this process. We propose a performance auto-tuning algorithm based on Collaborative Filtering and Auto Encoders in order to transfer useful knowledge between an already tuned source application and a novel target one. We test our method with datasets composed of data points collected with open-source benchmark suites and DBMS commonly employed in enterprise settings. We find that our algorithm is both fast and effective in finding new configurations to apply in the system to be tuned.