Workload Adaptations for Contended Main-Memory Multicore Transactions

Abstract

Database transactions guarantee atomicity for complex queries. Recent work in main- memory multicore transaction processing systems have achieved high transaction processing throughput, especially for uncontended workloads. Many systems achieve high performance by implementing new concurrency control (CC) protocols. Notably, variants of optimistic concurrency control (OCC), such as single-version concurrency control (1VCC) and multi- version concurrency control (MVCC) can perform well even under contention. In this work, I present MSTO, an MVCC system implemented to evaluate CC performance without the impact of basis factors [13]. Experimental results show that while MVCC does outperform over 1VCC in some scenarios, 1VCC is far more resilient to collapse at high contention than previously believed. 1VCC even outperforms MVCC on many high-contention workloads. I then introduce an optimization to reduce write-write conflicts between transactions, deferred updates. In conjunction with the static timestamp splitting [13] optimization that reduces read- write conflicts, deferred updates can be very effective at improving transactional throughput for all CC protocols, including TPC-C throughputs of 5.68× for 1VCC and 4.72× for MVCC compared to their baselines. Finally, I present and evaluate adaptive timestamp splitting, which changes each record’s partitioning strategy to accommodate workloads with heterogeneous access patterns. On workloads where 1VCC with deferred updates and adaptive timestamp splitting records the highest throughput, it commits up to 3.98× as many transactions as 1VCC with deferred updates and static timestamp splitting, and up to 4.43× as many transactions as baseline 1VCC. Not all workloads benefit from adaptive timestamp splitting, as baseline MVCC records greater throughput at very high contention.