Heterogeneous Data Access Model for Concurrency Control and Methods to Deal with High Data Contention
Alexander Thomasian
TL;DR
The paper investigates how concurrency control and lock contention impact OLTP performance under high-contention workloads, introducing a heterogeneous data access model (HDAM) to capture realistic access patterns. It combines HDAM with queueing-network theory to derive predictive formulas for load leading to thrashing and to analyze performance across single- and multi-class txn workloads and multiple DBRs. A key contribution is the quadratic relation for txn response time under contention, $a R^2(\lambda) - R(\lambda) + r(\lambda) = 0$, with thrashing onset at $\alpha^*=0.226$ and peak throughput near $\beta \approx 0.3$, along with a survey of restart-oriented CC methods (RP, WDL) that can outperform strict 2PL at additional CPU cost. The work emphasizes measurement-based parameterization and practical load-control guidance, suggesting how these models can inform performance prediction and CC strategy in contemporary systems (e.g., MySQL, Silo, MVCC/OCC variants) and prompting broader CC studies for diverse database architectures.
Abstract
OLTP has stringent performance requirements defined by Service Level Agreements. Transaction response time is used to determine the maximum throughout in benchmarks. Capacity planning tools for OLTP performance are based on queueing network models for hardware resources and database lock contention has a secondary effect on performance. With ever increasing levels of e-commerce and surges in OLTP traffic we discuss the need for studies of database workloads to develop more realistic lock/latch contention models. Predictive formulas to model increased load leading to thrashing for txns with identical and nonidentical steps are presented. We review concurrency control methods to reduce the level of lock/data conflicts in high contention environments.