Benchmarks

Netatalk Benchmarks

Benchmarks of file sharing server transfer speeds.

Comparative File Transfers

March 2024

The following benchmarks illustrate differences in file transfer speeds between Netatalk, Samba, and NFS.

Upload is a transfer from client to server. Download is a transfer from server to client.

• Server: Debian 12
• Client: macOS Sonoma
• Network: Both connected to the same Wi-Fi AP

Three data points per test case, with the median result is marked in bold below. Lower is better.

Server App	1x 2GB file Upload	Download	1000x 10KB files Upload	Download
Netatalk 3.1.18	42 45 46	44 45 48	84 86 94	50 55 58
Samba 4.17.12	45 47 52	45 48 52	99 100 101	53 53 54
NFS 2.6.2	204 205 208	136 139 139	97 97 99	24 24 25

Unit is seconds, rounded down to the integer

Results lantest version 0.x

January 2011

AFP Loadbalancing or Failover with multipe nodes each running Netatalk ontop of a cluster filesystem like GFS or OCFS.

The basic trick is to configure all Netatalk nodes to use one specific CNID server.

Some benchmark results follow below.

Test	Test details	Read 200 files	Write 30 MB	Read 30 MB	Lock/Unlock 4000x	Enumerate 320 files	Create ditree w. 10³ dirs
1	Linux x86_64, Xeon [email protected], FC RAID, ext3, localhost	0.235 s	0.116 s	0.053 s	0.472 s	0.426 s	0.548 s
2	Opensolaris x86_64, Phenom II [email protected] GHz, sATA, ZFS, localhost	0.047 s	0.061 s	0.040 s	0.139 s	3.334 s	11.885 s
3	HELIOS UB+ on Macpro x86 Xeon, builtin Harddisk, 100 MBit	0.301 s	3.077 s	2.957 s	3.997 s	0.014 s	1.587 s
4	Netatalk 2-node cluster on CentOS 5.4 x86 in Virtualbox VM on Opensolaris x86_64, Phenom II [email protected] GHz, iSCSI, GFS2, localhost	3.116 s	0.719 s	3.225 s	7.734 s	0.399 s	12.478 s
5	Same cluster as in 4), but this time with 2 cluster nodes running lantest concurrently	4.192 s	1.498 s	6.789 s	10.280 s	0.510 s	139.706 s
6	Netatalk 2-node cluster on Ubuntu 10beta1 x86 in Virtualbox VM on Opensolaris x86_64, Phenom II [email protected] GHz, iSCSI, OCFS2, localhost	0.170 s	0.780 s	0.583 s	0.839 s	0.111 s	1.336 s
7	Same cluster as in 6), but this time with 2 cluster nodes running lantest concurrently	0.490 s	0.858 s	1.366 s	1.646 s	0.276 s	41.288 s

Notes on individual tests

2) Prior to lantest version 1.0.0 this test time includes the time it took to create the test files. ZFS is just being safe here, honoring cache flush request from the CNID daemon. Currently the CNID daemon packs every CNID database modification into a dedicated database transaction which causes a cache flush. A fast ZIL may help.

4) A Netatalk Cluster using GFS2 as shared fs. Fast at some stuff, a little slow with other stuff. GFS2 is mounted with "noatime,nodiratime" options.

5) Apparently GFS2 doesn't like creating directories... But note: ALL traffic, including clusterwide locking stuff AND iSCSI had to pass in and out the VMs, so I'd expect a custer with real hardware to perform much better.

6) Another Netatalk cluster exactly speced like 4), but this time with OCFS2. This tests a very fresh version of OCFS2 which brings support for POSIX byterange locks. Without them Netatalk would be foobared. GFS2/OCFS2 shootout: OCFS2 wins hands down! Note: I had forgotten to set noatime mount option, using that would probably make things even faster.