A High-Performance Persistent Transactional Memory System via Cooperative Concurrency Control

Persistent Transactional Memory (PTM) simplifies crash-consistent programming on PM by providing atomicity and durability through transactional abstractions. Existing PTM systems suffer from fundamental limitations: sync-based PTMs couple transaction execution with data persistence, leading to poor concurrency and high latency. To address this, async-based PTMs decouple execution and persistence (log replay), but introduce significant coordination persistence overhead during log replay, which degrades performance. We present Courier, a high-performance PTM system that bridges execution and persistence through deterministic concurrency control. The key idea is to leverage the serializable transaction order established at runtime to deterministically schedule log replay, avoiding expensive coordination ordering overheads in persistence. To this end, Courier proposes a lightweight cooperative transaction framework to outline the persistence order in each individual log replay. By employing a series of techniques to coordinate with deterministic persistence order, Courier achieves effective and scalable log replay and avoids the need for sequential writes. Extensive evaluations using micro-benchmarks, YCSB, and TPC-C show that Courier improves throughput by up to 10.4 × and reduces 99th percentile latency by up to 16 × compared to state-of-the-art PTM systems.