Publish on Ping: A Better Way to Publish Reservations in Memory Reclamation for Concurrent Data Structures

TL;DR

This paper introduces a novel memory reclamation technique called publish-on-ping that reduces overhead in concurrent data structures by eliminating global reservations, leading to significant performance improvements over hazard pointers and hazard eras.

Contribution

The paper proposes a new memory reclamation method combining POSIX signals and delayed reclamation, serving as a drop-in replacement for hazard pointers and hazard eras, with enhanced performance and robustness.

Findings

Performance improvements of 1.2X to 4X over hazard pointers

Up to 20% faster than optimized hazard pointer implementations

Up to 3X faster than hazard eras

Abstract

Safe memory reclamation techniques that utilize per read reservations, such as hazard pointers, often cause significant overhead in traversals of linked concurrent data structures. This is primarily due to the need to announce a reservation, and fence to enforce appropriate ordering, before each read. In read-intensive workloads, this overhead is amplified because, even if relatively little memory reclamation actually occurs, the full overhead of reserving records is still incurred while traversing data structures. In this paper, we propose a novel memory reclamation technique by combining POSIX signals and delayed reclamation, introducing a publish-on-ping approach. This method eliminates the need to make reservations globally visible before use. Instead, threads privately track which records they are accessing, and share this information on demand with threads that intend to reclaim memory. The approach can serve as a drop-in replacement for hazard pointers and hazard eras. Furthermore, the capability to retain reservations during traversals in data structure operations and publish them on demand facilitates the construction of a variant of hazard pointers (EpochPOP). This variant uses epochs to approach the performance of epoch-based reclamation in the common case where threads are not frequently delayed (while retaining the robustness of hazard pointers). Our publish-on-ping implementations based on hazard pointers (HP) and hazard eras, when applied to various data structures, exhibit significant performance improvements. The improvements across various workloads and data structures range from 1.2X to 4X over the original HP, up to 20% compared to a heavily optimized HP implementation similar to the one in the Folly open-source library, and up to 3X faster than hazard eras. EpochPOP delivers performance similar to epoch-based reclamation while providing stronger guarantees.