Systems Seminar
April 18, 2012
1:00 p.m., AKW 200

Host: Daniel Abadi

Speaker: Andy Pavlo, Brown University
Title: Making Fast Databases Faster

Abstract: Anybody can make a fast database management system (DBMS) just by storing all of their data in main memory. The real challenge is in how one makes such systems go even faster and scale to support the demands of modern web- scale on-line transaction processing (OLTP) applications. Many of the so- called NoSQL systems are simply not an option for applications that are unable to relax their ACID requirements. Thus, a new emerging class of parallel main memory DBMSs, called NewSQL, are designed to take advantage of these application's partitionable workloads while maintaining traditional DBMS guarantees. But because storage I/O is no longer the bottleneck in a diskless environment, new challenges arise that often cannot be overcome just by adding more hardware. This talk will discuss our research in improving the performance of systems that are already fast to begin with. We will first present techniques for automatically partitioning a main memory, shared-nothing database in such a way that reduces the number of distributed transactions. We will then present a novel approach for dynamically selecting the proper transaction optimizations at run time using probabilistic models. Such optimizations are applied both before a transaction begins to execute, as well as while it executes.