Computer systems contain software which is becoming increasingly complex. Complex software often includes many configurable parameters or knobs that can be adjusted to adjust performance and energy consumption. As an example, a smartphone's settings include many on/off features and "sliders" one can adjust; but trying all the possible combinations to improve performance and reduce battery consumption is very time consuming. This project aims to optimize computer systems by (1) automatically exploring many parameter combinations and (2) helping humans see visual indications of how these parameters work and better understand complex systems.
This project will: (1) develop techniques to optimize storage systems, because they are the slowest part of any computer; (2) combine features from existing optimization and machine learning techniques; (3) improve the search for optimal settings by deciding when to stop and restart searching as well as considering the cost of changing system settings; (4) develop human driven visual techniques to explore extremely large sets of option combinations to better understand them and further direct the optimization process; and (5) evaluate all these techniques on real world storage systems.
Computer storage systems are so complex that no human can fully optimize them, particularly when circumstances change. This project will help automate the optimization of storage systems, improving their performance and energy use; advance the state of the art in hybrid optimization and machine learning techniques; develop and release interactive visualization systems that let humans understand, view, and direct a search process to promising directions; train and educate graduate and undergraduate students; and produce results that are applicable to other computer system optimization problems.
The project's artifacts such as software, source code, data sets, and results are part of a system called "Spectra". These artifacts will be made public through the project Website: www.filesystems.org/spectra. Results will be disseminated in peer-reviewed publications and on arxiv.org. The data will be maintained for at least ten years following the end of the project.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
