Integrating Non-Volatile Main Memory in a Deterministic Database

Deterministic databases provide strong serializability while avoiding concurrency-control related aborts by establishing a serial ordering of transactions before their execution. Recent work has shown that they can also handle skewed and contended workloads effectively. These properties are achieved by batching transactions in epochs and then executing the transactions within an epoch concurrently and deterministically. However, the predetermined serial ordering of transactions makes these databases more vulnerable to long-latency transactions. As a result, they have mainly been designed as main-memory databases, which limits the size of the datasets that can be supported. We show how to integrate non-volatile main memory (NVMM) into deterministic databases to support larger datasets at a lower cost per gigabyte and faster failure recovery. We describe a novel dual-version checkpointing scheme that takes advantage of deterministic execution, epoch-based processing and NVMM's byte addressability to avoid persisting all updates to NVMM. Our approach reduces NVMM accesses, provides better access locality, and reduces garbage collection costs, thus lowering the performance impact of using NVMM. We show that our design enables scaling the dataset size while reducing the impacts of using NVMM, achieving up to 79% of DRAM performance. Our design supports efficient failure recovery and outperforms alternative failure recovery designs, especially under contended workloads, by up to 56%.