Thursday, March 06, 2014, 11:00am
Ph.D Student, University of California Santa Barbara
"Elasticity Primitives for Database-as-a-Service"
Abstract: Database management systems (DBMSs) are the critical component in large-scale data-intensive applications. Traditional databases are predicated on an architecture that assumes the DBMS has full access to a single dedicated server. However, the current explosion in the amount of managed data has created a demand to virtualize databases into a database-as-a-service offering that is distributed across many shared-nothing servers. A database platform that consolidates many small databases (tenants), while dynamically load-balancing the placement of tenants calls for solutions that challenge traditional database systems. With hosted tenants subject to evolving workloads, ad-hoc usage, and flash crowds a data platform requires new primitives to place and load-balance the database without gross over-provisioning of resources.
This talk will review challenges in providing a database-as-a-service and elasticity primitives required to enable multitenancy using existing DBMS architectures. Pythia is presented as a self-managed system controller to model workload resource requirements, predict the impact of tenant colocation, and to resolve performance crisis that arise from resource starvation. Load-balancing mechanisms, such as Pythia, rely on the presence of lightweight migration primitives to move tenants between servers. This talk also presents Zephyr, a technique for the live migration of shared-nothing databases without incurring any downtime. These solutions are critical first steps in the building of a large scale, elastic, and self-managed database-as-a-service offering.
Bio: Aaron J. Elmore is a PhD candidate at the University of California Santa Barbara working with Divyakant Agrawal and Amr El Abbadi. Starting in April 2014, he will be a postdoctoral associate at MIT working with Samuel Madden and Michael Stonebreaker. His research interests include database virtualization, elastic database systems, data replication, and tools for collaborative scientific research.
Hosted by: EECS Prof. Gokhan Memik