A Performance Study of Locking Granularity in Shared-Nothing Parallel Database Systems

Locking granularity refers to the size of a lockable data unit, called a "granule", in a database system. Fine granularity improves system performance by increasing concurrency level but it also increases lock management overhead. Coarse granularity, on the other hand, sacrifices system performance but lowers the lock management cost. This paper explores the impact of granule size on performance of shared-nothing parallel database systems. We also report the interaction between transaction and sub-transaction scheduling policies and the locking granularity. Index Terms − Database systems, parallel systems, performance evaluation, locking, locking granularity, scheduling.