Persistence and Synchronization: Friends or Foes?

TL;DR

This paper evaluates how combining crash-consistency and synchronization techniques affects the performance and correctness of transactional systems on emerging non-volatile memory hardware, considering hardware features and workload traits.

Contribution

It provides a comprehensive analysis of different crash-consistency and synchronization methods for NVM transactional systems, highlighting tradeoffs and guiding design choices.

Findings

Persistence domain significantly impacts system performance.

Application transaction size influences overhead and efficiency.

Hardware features like HTM support affect implementation choices.

Abstract

Emerging non-volatile memory (NVM) technologies promise memory speed byte-addressable persistent storage with a load/store interface. However, programming applications to directly manipulate NVM data is complex and error-prone. Applications generally employ libraries that hide the low-level details of the hardware and provide a transactional programming model to achieve crash-consistency. Furthermore, applications continue to expect correctness during concurrent executions, achieved through the use of synchronization. To achieve this, applications seek well-known ACID guarantees. However, realizing this presents designers of transactional systems with a range of choices in how to combine several low-level techniques, given target hardware features and workload characteristics. In this paper, we provide a comprehensive evaluation of the impact of combining existing crash-consistency and synchronization methods for achieving performant and correct NVM transactional systems. We consider different hardware characteristics, in terms of support for hardware transactional memory (HTM) and the boundaries of the persistence domain (transient or persistent caches). By characterizing persistent transactional systems in terms of their properties, we make it possible to better understand the tradeoffs of different implementations and to arrive at better design choices for providing ACID guarantees. We use both real hardware with Intel Optane DC persistent memory and simulation to evaluate a persistent version of hardware transactional memory, a persistent version of software transactional memory, and undo/redo logging. Through our empirical study, we show two major factors that impact the cost of supporting persistence in transactional systems: the persistence domain (transient or persistent caches) and application characteristics, such as transaction size and parallelism.