Performance Optimization And Implementation Of Persistent File System For Non-volatile Memory

Emerging non-volatile main memories(NVMs)technologies such as PCM and 3D Xpoint are driving the development of computer storage architectures in recent years.They can provide both data persistence and high performance at memory level.Nonvolatile main memories offer higher density than SRAM,access performance and persistency like DRAM and disk respectively.It can also persist data at memory level.However,traditional file systems are block-based and designed for the disk access features,which fail to leverage the byte-addressable characteristic of NVMs efficiently.Consequently,there have been some NVMM file systems focusing on the study of data consistency in memory level.The design of existing file systems for NVM has to handle the data durability problem between CPU cache and NVM.However,most NVMM file systems could not meet the strong data consistency requirement of applications with data structures.The contributions of this work are summarized as below.(1)Comparing several mainstream traditional file systems with state-of-the-art NVMM file systems,analysis and evaluationWe use DRAM to emulate NVM,evaluate these file systems with specific workloads and analyze the experimental results to find out which structural features can play a key role in improving the performance of the file system.From the experimental results,we can see that some of the traditional file system strategies cannot make the most advantage of the new media.In spite of moving it to the memory level,the performance gap with the NVMM file system still exists.(2)Proposing an optimization strategy for small data write and designing a variety of experiments to verify and evaluate the resultsBy analyzing the performance of each file system,the bottleneck of the logstructured file system for non-volatile memory in small data workload was found and optimized,to provide high performance in small data-intensive applications.We use inplace writes to minimize memory footprint for small data updating and accelerate data block locating with hashing strategy.The experimental results show that this improves the system throughput by up to 23 times compared to state-of-the-art file systems and reduces the execution time by up to 65.5%.(3)Proposing a data consistency mechanism to provide strong data consistency guarantees of both metadata and data.We conducted a detailed analysis and comparison of commonly used data consistency mechanisms.Based on these methods,we adopt a small data log to journal fine-grained data writes.Combining this with existing mechanisms can maximize the benefits of NVM with the assurance of data consistency.