Learned Cost-Models for Robust Tuning

Note: Please see the Robust Data Systems Tuning project page for earlier results associated with this research.

Abstract: Data systems’ performance is tuned via analytical cost models that take into account all tuning knobs and predict performance. However, as the complexity of data systems increases, there are more tuning knobs and, as a result, analytical cost models become cumbersome to use and, in some cases, impossible to derive. We propose to augment cost-based tuning with Learned Cost Models that (i) can be trained using analytical cost models to allow for more flexible tuning and (ii) can learn from past and (targeted) future executions to capture the increased system complexity.

This project is supported by the Red Hat Collaboratory at Boston University.


Research Area(s)

Project Resources