III: Small: Database Processing on GPUs

Modern Graphic Processing Units (GPUs) offer more parallelism and higher memory bandwidth than CPUs. This project aims to take advantage of these properties by developing a system to efficiently process database queries over GPU-resident datasets. To achieve this goal this project employs the following approaches: (a) The development of novel indexing techniques that combine multidimensional partitioning with block-oriented bitmaps, and whose parameters are sensitive to the query distribution;(b) The optimization of memory bank contention and value contention between threads; (c) The efficient implementation of a complete set of relational database operators, including aggregation, joins, and indexed selections; and (d) The evaluation of the performance of the system on query-intensive workloads, using real applications and standard benchmarks. Improvements in database system performance would have wide-ranging impact on the efficiency of many enterprises that employ database systems for analytics. The project supports PhD students working on database system implementation techniques. The innovations and software created during the project will be used to enhance the curriculum of the Database Systems Implementation course at Columbia University. Publications, software, and other project data will be disseminated via the web at our project web site (http://www.cs.columbia.edu/~kar/gpuproject.html).