An Evaluation of Strict Timestamp Ordering Concurrency Control for Main-Memory Database Systems

Abstract

With the fundamental change of hardware technology, main-memory database systems have emerged as the next generation of DBMS. Thus, new methods to execute transactions in a serial, lock-free mode have been investigated and successfully employed, for instance in H-Store or HyPer. Although these techniques allow for unprecedentedly high throughput for suitable workloads, their throughput quickly diminishes once unsuitable transactions, for instance those crossing partition borders, are encountered. Still, little research concentrates on the overdue re-evaluation of traditional techniques, that do not rely on partitioning.This paper studies strict timestamp ordering (STO), a "good old" technique, in the context of modern main-memory database systems built on commodity hardware with high memory capacities. We show that its traditional main drawback -slowing down reads -has a much lower impact in a main-memory setting than in traditional diskbased DBMS. As a result, STO is a competitive concurrency control method which outperforms the partitioned execution approach, for example in the TPC-C benchmark, as soon as a certain percentage of the workload crosses partition boundaries.

Exploiting hardware transactional memory in main-memory databases

“…In an earlier evaluation we showed that timestamp-based concurrency control has become a promising alternative to traditional locking [35]. Lomet et al [36] and Larson et al [8] recently devised multi-version concurrency control schemes that, like our approach, use a timestamp-based version control to determine conflicting operations.…”

Scaling HTM-Supported Database Transactions to Many Cores

“…They propose a more lightweight scheme, where, instead of locks in a global lock manager data structure, each tuple has two counters that indicate how many transactions requested read or write access. In an earlier evaluation we showed that timestamp-based concurrency control has become a promising alternative to traditional locking [34].…”

In-Memory Big Data Management and Processing: A Survey

“…Very Lightweight Locking (VLL) [106] further simplifies the data structure by compressing all the lock states of one record into a pair of integers for partitioned databases. Another class of concurrency control is based on timestamp, where a predefined order is used to guarantee transactions' serializability [144], such as optimistic concurrency control (OCC) [39], [107] and multi-version concurrency control (MVCC) [108], [109]. Furthermore, H-Store [36], [101], seeks to eliminate concurrency control in single-partition transactions by partitioning the database beforehand based on a priori workload and providing one thread for each partition.…”