Persistent Memory Programming Abstractions in Context of Concurrent Applications

TL;DR

This paper evaluates different programming abstractions for non-volatile memory in concurrent applications, analyzing their tradeoffs in ease of programming, performance, and consistency through practical data structure implementations.

Contribution

It provides a comparative analysis of NVthreads, Mnemosyne, and NVML abstractions by implementing concurrent persistent data structures and assessing their tradeoffs.

Findings

NVML is easiest to program but less efficient.

Mnemosyne offers better performance but is more complex to develop with.

Tradeoffs involve balancing ease of programming, efficiency, and consistency.

Abstract

The advent of non-volatile memory (NVM) technologies like PCM, STT, memristors and Fe-RAM is believed to enhance the system performance by getting rid of the traditional memory hierarchy by reducing the gap between memory and storage. This memory technology is considered to have the performance like that of DRAM and persistence like that of disks. Thus, it would also provide significant performance benefits for big data applications by allowing in-memory processing of large data with the lowest latency to persistence. Leveraging the performance benefits of this memory-centric computing technology through traditional memory programming is not trivial and the challenges aggravate for parallel/concurrent applications. To this end, several programming abstractions have been proposed like NVthreads, Mnemosyne and intel's NVML. However, deciding upon a programming abstraction which is easier to program and at the same time ensures the consistency and balances various software and architectural trade-offs is openly debatable and active area of research for NVM community. We study the NVthreads, Mnemosyne and NVML libraries by building a concurrent and persistent set and open addressed hash-table data structure application. In this process, we explore and report various tradeoffs and hidden costs involved in building concurrent applications for persistence in terms of achieving efficiency, consistency and ease of programming with these NVM programming abstractions. Eventually, we evaluate the performance of the set and hash-table data structure applications. We observe that NVML is easiest to program with but is least efficient and Mnemosyne is most performance friendly but involves significant programming efforts to build concurrent and persistent applications.