Laser: Load-adaptive group commit in lock-free transaction logging

Abstract

Log manager is a key component of DBMS and is considered as the most prominent bottleneck in the modern in-memory OLTP system. In this paper, by addressing two existing performance hurdles in the current procedure, we propose a high-performance transaction logging engine Laser and integrate it into OceanBase, an in-memory OLTP system. First, we present a lock-free transaction logging framework to eliminate the lock contention. Then we make theoretical analysis and propose a judicious grouping strategy to determine an optimized group time for different workloads. Experiment results show that it improves 1.4X–2.4X throughput and reduces more than 60% latency compared with current methods.