Experiment-driven System Management

August 5, 2010
Shivnath Babu | Duke University

Experiment-driven management is an important and overlooked part of system administration that aims to automate the trial-and-error practices used to perform certain administrative tasks today. I will first present how experiment-driven management simplifies tuning, benchmarking, testing, troubleshooting, and capacity planning for a variety of systems like databases, storage servers, and MapReduce frameworks. I will then present Xplus, a new query optimizer that employs principles of experiment-driven management to tune poorly-performing SQL queries. A user or higher-level tuning tool can tell Xplus: “I am not satisfied with the performance of the plan being used for the query Q that runs repeatedly. Can you generate a 50% better plan?” Xplus will then select a small set of subplans to run for Q, iterate based on the monitoring information collected, and either return a 50% better plan or a guarantee that no such plan exists for the current database configuration and optimizer cost model. Xplus uses a new abstraction of the physical plan space and an extensible architecture comprising multiple SQL-tuning experts. Finally, I will discuss why it is important to make experiment-driven management accessible to a wide range of users beyond system administrators; and present ongoing work on a language and run-time system being developed to achieve this goal on cloud computing platforms.

Speaker Details

Shivnath Babu is an Assistant Professor of Computer Science at Duke University. He got his Ph.D. from Stanford University in 2005. He has received a U.S. National Science Foundation CAREER Award and three IBM Faculty Awards. His current research focuses on making data-intensive computing systems easier to manage.