Empirical Study of Transactional Management for Persistent Memory

Emerging Non-Volatile Memories (NVMs) possess unique features including byte-addressability, non-volatility and DRAM-like performance, while consistency and persistency are the two main concerns when using NVM. Transaction persistent memory mechanism is proposed and widely-used for its fast and atomic operation to achieve persistency and consistency simultaneously in NVM. However, the transaction cost for NVM is comparatively huge and the reasons are still puzzles. In this paper, we empirically study the performance cost of NVM transaction by using transaction-related part of NVML library from Intel. We carry out extensive experiments and analyze the results in-depth to find out some internal reasons which impact the overall performance. Our experimental results show that the main performance bottlenecks locate at the transaction execution and commit processes, and the performance is also impacted heavily by adopting different memory allocation and update strategies. Finally, we give out some hints and recommendations for improving the design and implementation of NVML library and aiming to boost up the transactional performance in NVML and other persistent memory systems.