The emergence of large main memories and massively parallel processors has triggered the development of multi-core main-memory database management systems (DBMSs). Although the reduction of disk accesses results in low single-thread transaction execution time, scaling these systems on multi-core machines remains notoriously difficult. In particular, the concurrent processing of a large number of transactions can bring about significant performance bottlenecks.

In this talk, I will discuss the potential for improving the DBMS performance through program analysis. The intuition is that a DBMS can extract hidden parallelization opportunities by understanding the dependencies among operations. I propose two program-analysis-assisted techniques, transaction healing and PACMAN, which respectively attempts to optimize the performance of transaction processing and failure recovery. Our extensive experimental studies demonstrate that these two mechanisms can enable modern DBMSs to scale towards dozens of cores when supporting various types of transactional workloads.

Bio:
Yingjun Wu is a fifth year Ph.D. student from National University of Singapore. His research focuses on the design and implementation of transactional main-memory database systems.