Migrate to NFS Azure File Shares from Linux

Summary — Migrate to NFS Azure file shares

Overview

This article explains how to migrate data from Linux file servers to NFS Azure file shares (available only as Premium file shares on SSD).
It compares single-threaded rsync with multithreaded fpsync (which uses rsync/cpio/tar under the hood) and gives guidance for commonly used migration phases (baseline, incremental, final pass).
Key note: Azure Files doesn't support NFS access control lists (ACLs).

Applies to

NFS access is supported only on Premium file shares (FileStorage) with LRS/ZRS. Standard GPv2 file shares do not support NFS.

Prerequisites

Mount at least one NFS Azure file share on a Linux VM. See: https://docs.azure.cn/en-us/storage/files/create-classic-file-share
Recommended: mount with nconnect (e.g., nconnect=8) to use multiple TCP connections for better performance. See: https://docs.azure.cn/en-us/storage/files/nfs-performance#nfs-nconnect

Tool selection and behavior

rsync: single-threaded, flexible, widely used.
fpsync: multithreaded wrapper that partitions the source (via fpart) and runs parallel synchronization jobs; can use rsync (default), cpio, or tar as the copy tool.
- Important: fpsync synchronizes the contents of the source directory (enforces trailing '/'); it does not create a parent folder named after the source in the destination.
For large distributed filesystems, reducing per-file network round-trips and parallelizing transfers improves throughput.

Install fpart (required for fpsync)

Ubuntu:
```
sudo apt-get install fpart
```
RHEL 7/8/9 (example for enabling EPEL and installing fpart):
```
sudo yum install -y https://dl.fedoraproject.org/pub/epel/epel-release-latest-7.noarch.rpm
sudo yum install fpart -y
```
(use the appropriate EPEL URL for RHEL 8/9 as shown in the original guide)
If no package is available, install fpart from source: https://www.fpart.org/#installing-from-source

Migration workflow (three phases)

Baseline copy

First full copy when destination is empty.
Recommended tool/flags: fpsync with cpio.

Command form:

fpsync baseline example

fpsync -m cpio -n <parallel transfers> <absolute source path> <absolute destination path>

Incremental copy

Repeated runs to capture changes made since the baseline copy.
Recommended: fpsync with default copy tool (rsync). Run multiple times to converge.

Command form:

fpsync incremental example

fpsync -n <parallel transfers> <absolute source path> <absolute destination path>

Final pass

Remove files on destination that no longer exist on the source.
Options:
- Use rsync directly with --delete to reconcile: rsync --delete ...
- Or use fpsync with the -E option (see fpsync docs).

Typical fpsync invocation

fpsync -m <rsync|cpio|tar> -n <parallel transfers> <absolute source path> <absolute destination path>

Performance findings (high-level)

Multithreaded fpsync substantially improved throughput and IOPS versus single-threaded rsync in the authors' tests.
Distributing files across directories (so work can be parallelized) improves performance.
Larger file sizes yield better throughput than many small files.

Selected numeric summary (from the article tests)

Config

Files

Dirs

Size per file

Total size

rsync throughput

fpsync throughput

Throughput gain

1.1 (baseline)

1,000,000

0–32 KiB

18 GiB

0.33 MiB/s

1.20 MiB/s

267%

1.2 (incremental)

1,000,000

0–32 KiB

18 GiB

3.25 MiB/s

36.41 MiB/s

1,020%

2 (baseline)

191,345

3,906

0–32 KiB

3 GiB

0.27 MiB/s

6.04 MiB/s

2,164%

3 (baseline)

5,000

10 MiB

50 GiB

105.04 MiB/s

308.90 MiB/s

194%

Notes and caveats

Tests used Standard_D8s_v3 VMs (8 vCPUs, 32 GiB RAM) and NFS Azure File shares with >1 TiB provisioned.
Best observed fpsync settings in these tests: 64 threads when using rsync and 16 threads when using cpio with nconnect=8 on the mount. Actual results depend on your dataset and environment.
Some experiments were intentionally small; Azure Files throughput can be higher in other configurations.
Third-party tools (fpsync, fpart, rsync, cpio) are not Microsoft-supported; review licenses and support yourself.