Disk-based database systems benefit from concurrency among transactions ― usually with marginal overhead. For main-memory database systems, however, locking overhead can have a serious impact on performance. This paper proposes SP, a serial protocol for the execution of transactions in main-memory systems, and evaluates its performance against that of strict two-phase locking. The novelty of SP lies in the use of timestamps and mutexes to allow one transaction to begin before its predecessors' commit records have been written to disk, while also ensuring that no committed transactions read uncommitted data. We demonstrate seven-fold and two-fold increases in maximum throughput for read-and update-intensive workloads, respectively. At fixed loads, we demonstrate ten-fold and two-fold improvements in response time for the same transaction mixes. We show that for a wide range of practical workloads, SP on a single processor outperforms locking on a multiprocessor, and then present a modified SP, that exploits multiprocessor systems.
展开▼
