Database management systems (DBMSs) are ubiquitous in industry and their performance and functionality are crucially important in commercial IT infrastructure. There has been a lot of work on improving the performance of these critical systems. This project adopts a complementary and orthogonal approach to substantially increasing the performance of DBMSs. This is done by exploiting runtime information about the DBMS to specialize the DBMS code on the fly in order to eliminate as much unnecessary work as possible. This dynamic code optimization, which we term "micro-specialization", is highly aggressive and goes well beyond what can be achieved using existing compiler optimization technology. This focus on runtime code optimization is also very different from most of the current research on DBMS performance improvement. The project investigates algorithms for automatic identification of code sequences where micro-specialization may profitably be applied; algorithms that can micro-specialize such identified target code sequences; and techniques for ensuring the correctness of the resulting code. Preliminary studies on disparate DBMSes suggest that micro-specialization offers significant performance improvements when applied to mature, high-performance DBMSs and to industry-standard benchmarks, and that there are rich opportunities for other substantial performance gains, with minimal impact on the DBMS architecture.
Broader impacts of this project include increased efficiency of the IT infrastructure used by a wide variety of companies, leading to improved overall productivity. Graduate and undergraduate students are involved in all aspects of the research activities as an integral part of the project. The PIs have an established track record of integrating research activities into the undergraduate curriculum and involving undergraduates, including women and members of underrepresented minorities, in research; this project continues this tradition.
For further information see the web site at www.cs.arizona.edu/projects/microspecialization/ .
