EE Seminar: Accelerating DBMS on modern computer architecture using Persistent Memory
Speaker: Netanel Katzburg,
M.Sc. student under the supervision of Prof. Shlomo Weiss and Dr. Amit Golander
Wednesday, July 12th, 2017 at 15:30
Room 011, Kitot Bldg., Faculty of Engineering
Accelerating DBMS on modern computer architecture using Persistent Memory
Abstract
The architecture of Database Management Systems (DBMS) is closely related to the characteristics of the storage hierarchy, because durability and response time are highly dependent on the physical properties of the target storage. Main memory volatility requires a DBMS to provide durability by software means as data continuously moves between volatile memory buffers and input/output persistent media. Traditional storage systems applications use complex concurrency control schemes to reduce latency and increase throughput and in order to utilize multicore hardware and shared system resources. New persistent memory (PM) devices emerging in the last decade, such as PCM, RRAM and MRAM, exhibit near-DRAM speed and characteristics, provide data persistence, and could be game changing for storage bound applications. I demonstrate that novel file systems based on PM offer outstanding performance. As a result of this 280x speedup, for many applications PM-based file systems are likely to replace traditional file systems. With new PM-based file systems, the performance bottleneck of traditionally disk-bounded applications moves to the application code itself. Traditional system level tradeoffs are no longer significant and any further optimizations should be done at the application level. I focus on benefits of persistent memory and their impact on database management systems and consider methods for application speedup that are applicable to DBMSs that use PM. These optimization methods depend on the characteristics of PM storage. I consider concurrency and mutual resource contention, explore and rethink major application components, and finally combine static code optimization. Running the on-line transaction processing (OLTP) workload, the Relational Database Management Systems (RDBMSs) explored here show performance gains relative to traditional storage systems by a factor of 3.17 and 1.79 for PostgreSQL and SQLite respectively.
