Practical Detectability for Persistent Lock-Free Data Structures

TL;DR

This paper introduces a practical framework for persistent lock-free data structures with detectability, enabling fault-tolerance and exactly-once semantics in persistent memory systems, while maintaining high performance.

Contribution

It presents a novel programming framework supporting primitive detectable operations, systematic transformation of DRAM data structures to persistent memory, and crash recovery, advancing persistent lock-free data structures.

Findings

Implementations of queues and hash tables with detectability outperform existing non-detectable counterparts.

The framework enables crash recovery from full system and thread crashes effectively.

Performance is comparable to non-detectable data structures due to scratchpad optimization.

Abstract

Persistent memory (PM) is an emerging class of storage technology that combines the benefits of DRAM and SSD. This characteristic inspires research on persistent objects in PM with fine-grained concurrency control. Among such objects, persistent lock-free data structures (DSs) are particularly interesting thanks to their efficiency and scalability. One of the most widely used correctness criteria for persistent lock-free DSs is durable linearizability (Izraelevitz et. al., DISC 2016). However, durable linearizability is insufficient to use persistent DSs for fault-tolerant systems requiring exactly-once semantics for storage systems, because we may not be able to detect whether an operation is performed when a crash occurs. We present a practical programming framework for persistent lock-free DSs with detectability. In contrast to the prior work on such DSs, our framework supports (1) primitive detectable operations such as space-efficient compare-and-swap, insertion, and deletion; (2) systematic transformation of lock-free DSs in DRAM into those in PM requiring modest efforts; (3) comparable performance with non-detectable DSs by DRAM scratchpad optimization; and (4) recovery from both full system and thread crashes. The key idea is memento objects serving as a lightweight, precise, and per-thread checkpoints in PM. As a case study, we implement lock-free and combining queues and hash tables with detectability that outperform (and perform comparably) the state-of-the-art DSs with (and without, respectively) detectability.