Robust Data Systems Tuning
Note: Please see the Learned Cost-Models for Robust Tuning project page for research that has grown from this project. See the Robust LSM-Trees Under Workload Uncertainty project page for earlier results associated with this research.
BU faculty members Manos Athanassoulis and Evimaria Terzi will work on building a new robust tuning framework for LSM-based data systems that will allow to deploy and effectively tune data systems even when the workload information is inaccurate or noisy. This approach has the potential to revolutionize tuning and deployment of cloud-based data systems, which face volatility in resources availability and workload as multiple applications are collocated on shared virtualized infrastructure. The proposed paradigm addressed the observed workload uncertainty by (re-)formulating the tuning problem as a min-max optimization problem that seeks to find the best tuning for the worst-case workload within an uncertainty region. “The proposed robust tuning paradigm will also improve the collective understanding of robust tuning for data systems in general, a useful tool as new applications face increased workload volatility,” the team wrote.
- Gave tutorial at SIGMOD 2022 covering LSM trees which included a short discussion about our VLDB results, title “Dissecting, Designing, and Optimizing LSM-based Data Stores”
- Awarded NSF CAREER on “Robust LSM Trees,” part of which builds on the robust tuning started with this project
- May 2021 – Greater New England RIG Monthly Meeting “Robust LSM-Trees Tuning for Workload and Resources Uncertainty” by Andy Huynh
- May 2022, Meetup: Log-structured Merge (LSM) Trees in the Cloud Era
- Aug 2022, Meta invited talk at RocksDB team, “Endure: A Robust Tuning Paradigm for LSM Trees Under Workload Uncertainty” by Andy Huynh
- Sep 2022, At the VLDB 2022 conference, “Endure: A Robust Tuning Paradigm for LSM Trees Under Workload Uncertainty” by Andy Huynh