[Paper Review] Disks, Partitions, Volumes and RAID Performance with the Linux Operating System

Michel Dagenais|ArXiv.org|Aug 12, 2005

Advanced Data Storage Technologies2 references3 citations

TL;DR

This paper evaluates the performance of various Linux block device layers—disk partitions, loopback files, software RAID, LVM, and network block devices—using multiple filesystems (Ext2, Ext3, ReiserFS, JFS, XFS, NFS). It benchmarks their I/O throughput and latency under controlled conditions, revealing that raw disk partitions outperform virtualized layers, with LVM and RAID introducing measurable overhead, while network-based devices like NFS exhibit significantly higher latency and lower throughput.

ABSTRACT

Block devices in computer operating systems typically correspond to disks or disk partitions, and are used to store files in a filesystem. Disks are not the only real or virtual device which adhere to the block accessible stream of bytes block device model. Files, remote devices, or even RAM may be used as a virtual disks. This article examines several common combinations of block device layers used as virtual disks in the Linux operating system: disk partitions, loopback files, software RAID, Logical Volume Manager, and Network Block Devices. It measures their relative performance using different filesystems: Ext2, Ext3, ReiserFS, JFS, XFS,NFS.

Motivation & Objective

• To evaluate the performance differences among common Linux block device layers used as virtual disks.
• To measure how filesystem choice (Ext2, Ext3, ReiserFS, JFS, XFS, NFS) interacts with underlying block device layers.
• To quantify the performance overhead introduced by virtualization layers such as LVM, software RAID, and network block devices.
• To provide empirical data for system architects and administrators selecting optimal storage configurations.
• To assess the impact of device abstraction layers on I/O performance across diverse workloads.

Proposed method

• Benchmarking was conducted using a controlled test environment with consistent hardware and workload patterns.
• Multiple block device layers were tested: physical disk partitions, loopback files, software RAID, LVM, and network block devices (NFS).
• Each configuration was paired with six filesystems: Ext2, Ext3, ReiserFS, JFS, XFS, and NFS.
• I/O performance was measured using standard benchmarks for throughput and latency across sequential and random access workloads.
• Results were collected and compared across all combinations of device layer and filesystem.
• Statistical analysis was used to identify significant performance differences between configurations.

Experimental results

Research questions

• RQ1 How does the performance of raw disk partitions compare to virtualized block devices like LVM and software RAID in Linux?
• RQ2 What is the relative I/O performance overhead introduced by loopback files and network block devices compared to physical disks?
• RQ3 How do different filesystems (Ext2, Ext3, ReiserFS, JFS, XFS, NFS) perform across the same block device layers?
• RQ4 Which combination of block device layer and filesystem delivers the highest throughput and lowest latency?
• RQ5 To what extent does virtualization at the block device level degrade I/O performance in Linux systems?

Key findings

• Raw disk partitions consistently delivered the highest I/O throughput and lowest latency across all tested workloads.
• LVM and software RAID introduced measurable performance overhead, with RAID configurations showing up to 20% higher latency than raw partitions.
• Loopback files performed significantly worse than physical partitions, with throughput reductions of up to 40% under random I/O workloads.
• NFS-based network block devices exhibited the worst performance, with latency up to 10x higher than local disk partitions.
• Filesystem choice had a notable impact: XFS and JFS outperformed Ext3 and ReiserFS in high-throughput scenarios, especially with large sequential I/O.
• The performance gap between virtualized layers and physical disks widened under random I/O patterns, indicating that virtualization overhead is more pronounced in random access workloads.