Computational Storage: Make Relational Databases Efficiently Support Data Analytics

Recent years have witnessed a significant interest in the development of a single database system that can efficiently handle both transactional and analytical workloads to support the demands of real-time business decision making. This talk will present a solution that employs emerging computational storage technology to make traditional transaction-oriented database cost-effectively support analytical workloads. The key idea is to pushdown certain expensive analytics-centric tasks (e.g., table scan) from CPUs to the underlying computational storage devices. Practical implementation of this simple idea is nontrivial and demands cohesive design innovations across the software (i.e., database and I/O stack) and hardware (i.e., computational storage device) layers. This talk will present such a holistic implementation for MySQL, the most widely used open-source database. Its practical effectiveness has been successfully demonstrated through extensive TPC-H benchmarking.