HiNFS：非易失性主存储器的高性能文件系统设计

79 浏览量更新于2024-07-14 收藏 537KB PDF 举报

"这篇研究论文探讨了非易失性主存储器(NVMM)的高性能文件系统设计，提出了一种名为HiNFS的新方案，旨在解决NVMM技术中写入速度慢的问题，通过优化写入缓冲策略提升系统性能。" 在计算机科学领域，新兴的非易失性主内存技术（NVMM）正在改变存储架构，它可以在主存级别提供数据持久性，从而无需传统的磁盘或闪存存储。然而，现有的NVMM意识的文件系统为了避免用户缓冲区、操作系统页缓存和存储层之间的双倍复制开销，通常会绕过OS页缓存，直接在用户缓冲区和NVMM存储之间复制数据。尽管这减少了中间层的复杂性，但NVMM的写入速度较慢成为了性能瓶颈。论文作者针对这一问题提出了HiNFS（HighPerformanceFileSystem for Non-Volatile Main Memory），一个专为NVMM设计的高性能文件系统。HiNFS的核心创新在于引入了一个NVMM感知的WriteBuffer策略。这个策略允许文件系统的写入操作在DRAM（动态随机存取内存）中缓冲延迟持久化的文件写入，并且延迟地将这些数据持久化到NVMM，以此来隐藏NVMM的长写入延迟。通过这种方式，HiNFS能够在不影响数据安全性的前提下，提高系统的整体性能。同时，HiNFS对于那些需要立即持久化的操作，如系统关键数据的更新，仍然执行直接访问NVMM的方式，以确保这些数据的即时性和可靠性。这种方法结合了延迟写入和直接访问的优势，为NVMM环境下的文件系统操作提供了更为灵活和高效的解决方案。此外，HiNFS可能还需要考虑其他因素，比如错误恢复机制、数据一致性保证以及 NVMM 的寿命管理等。论文可能会深入讨论这些方面，如何在HiNFS的设计中平衡性能、可靠性和耐用性，以适应不同应用场景的需求。这篇研究论文为解决NVMM的性能问题提供了一种创新的文件系统设计方案，通过优化写入策略，提升了在非易失性主内存环境下运行的文件系统的整体效能。这对于未来数据中心和云计算环境中的高速存储系统设计具有重要的参考价值。

User space

Kernel

File IO

Application

VFS

Traditional

DRAM

Generic Block Layer

NVMM

(a) Traditional Block-based File System

User space

Kernel

File IO

Application

VFS

NVMM-FS

NVMM

(b) NVMM-aware File System

User space

Kernel

File IO

Application

VFS

HiNFS

DRAM

NVMM

Read &&

Lazy-Persistent

Write

Read &&

Eager-Persistent

Write

memory interface

Figure 3. Architecture Comparison of Different File Systems for NVMM.

LASR has no fsync writes. To conclude, a large number

of applications have a signiﬁcant portion of lazy-persistent

writes, which are consistent with prior research results [19].

The above observations have interesting implications for

the design of the ﬁle system for fast NVMM. On one hand,

the revealed direct write access overhead strongly suggests

that we need to reduce prompt writes to NVMM in order to

improve the performance. On the other hand, we believe that

an elegant design should be ﬂexible. In other words, it should

not improve the performance in some particular cases, while

sacriﬁcing the performance in other cases. For example,

simply using DRAM as a cache of NVMM may improve

the performance for workloads having many lazy-persistent

writes, but this simple design will signiﬁcantly degrade the

system performance for workloads containing many eager-

persistent writes due to the double-copy overheads.

3. HiNFS Design

In this section, we ﬁrst describe the high-level system ar-

chitecture comparison of existing ﬁle systems and HiNF-

S. We then present an NVMM-aware Write Buffer policy

to reduce prompt writes to NVMM by buffering the lazy-

persistent writes in DRAM temporarily. Finally, we discuss

how to eliminate the double-copy overheads resulted from

conventional buffer management.

3.1 System Architecture

Figure 3(a) shows the system architecture of traditional

block-based ﬁle systems on a RAMDISK-like NVMM block

device. This is the most straightforward way to use NVMM

as a persistent storage in which legacy ﬁle systems, such as

ext2/ext4, can directly work on NVMM without extra modi-

ﬁcations by emulating it as a block device. In a block-based

ﬁle system, each ﬁle I/O usually requires two data copies,

one between the block device and the OS page cache through

the generic block layer, and one between the OS page cache

and the user buffer through the memory interface. Howev-

er, it has been recently reported that the overheads from the

double-copy and the generic block layer can signiﬁcantly

impact the NVMM system performance [6, 13, 18, 49]. As

a result, state-of-the-art NVMM-aware ﬁle systems, such as

BPFS [13], PMFS [18], etc., access the NVMM device di-

rectly as shown in Figure 3(b). In these NVMM-aware ﬁle

systems, each ﬁle I/O requires only a single data copy, di-

rectly between the NVMM and the user buffer (a.k.a. direct

access). Unfortunately, the major drawback of this approach

is that it does not consider NVMM’s relatively longer write

latency compared to DRAM. Speciﬁcally, each write opera-

tion leads to prompt access to NVMM, which always expose

the long write latency of NVMM to the critical path, lead-

ing to suboptimal system performance. Therefore, to get the

best system performance, we propose another system archi-

tecture for the NVMM storage as shown in Figure 3(c). The

design objectives of HiNFS are twofold:

(1) Hiding the long write latency of NVMM behind the crit-

ical path. HiNFS uses an NVMM-aware Write Buffer

policy to buffer the lazy-persistent writes in DRAM tem-

porarily. HiNFS design, including ﬁne-grained buffer

management and using a memory interface to interact

between DRAM and NVMM, is optimized for the N-

VMM storage. (Section 3.2)

(2) Eliminating the double-copy overheads. Although

buffering can help hide the long write latency of NVM-

M, it may introduce the double-copy overheads. For this

reason, HiNFS optimizes read and eager-persistent write

by avoiding unnecessary data copies. Read or eager-

persistent write, in HiNFS, requires only a single data

copy between DRAM/NVMM and the user buffer. (Sec-

tion 3.3)

3.2 NVMM-aware Write Buffer Policy

To hide the relatively long write latency of NVMM be-

hind the critical path, we propose an NVMM-aware Write

Buffer policy to buffer the lazy-persistent writes in DRAM

temporarily. Figure 4 shows an overview of HiNFS. When a

write request is serviced, the Eager-Persistent Write Check-

er module would decide whether the current write operation

is a lazy-persistent or eager-persistent write. We would like

to discuss the approach of identifying the eager-persistent

剩余15页未读，继续阅读

weixin_38531788

粉丝: 4
资源: 913

HiNFS：非易失性主存储器的高性能文件系统设计

文件存储系统

开源文件系统,超高性能

NVMTFS：用于分层存储系统的非易失性存储器自适应文件系统

电信设备-非易失性存储器阵列中的保存单元信息的非易失性存储器.zip

行业文档-设计装置-用于运行串行非易失性半导体存储器的方法.zip

非易失性存储器数据掉电保护的硬件解决方案

linux-nova:NOVA是一种日志结构的文件系统，专门为字节寻址的非易失性存储器而设计，由加利福尼亚大学圣地亚哥分校开发

NVMTFS：非易失性存储器适应的分层存储文件系统

非易失性存储器写入机制深度解析

探索非易失性存储器平台的可配置分区技术

最新资源