Latch-free data structures for DBMS: design, implementation, and evaluation

The fact that multi-core CPUs have become so common and that the number of CPU cores in one chip has continued to rise means that a server machine can easily contain an extremely high number of CPU cores. The CPU scalability of IT systems is thus attracting a considerable amount of research attention. Some systems, such as ACID-compliant DBMSs, are said to be difficult to scale, probably due to the mutual exclusion required to ensure data consistency. Possible countermeasures include latch-free (LF) data structures, an elemental technology to improve the CPU scalability by eliminating the need for mutual exclusion. This paper investigates these LF data structures with a particular focus on their applicability and effectiveness. Some existing LF data structures (such as LF hash tables) have been adapted to PostgreSQL, one of the most popular open-source DBMSs. The performance improvement was evaluated with a benchmark program simulating real-world transactions. Measurement results obtained from state-of-the-art 80-core machines demonstrated that the LF data structures were effective for performance improvement in a many-core situation in which DBT-1 throughput increased by about 2.5 times. Although the poor performance of the original DBMS was due to a severe latch-related bottleneck and can be improved by parameter tuning, it is of practical importance that LF data structures provided performance improvement without deep understanding of the target system behavior that is necessary for the parameter tuning.