The ultimate Linux filesystems benchmark

Having a car accident makes you have a lot of free time.

So I decided to create the ultimate Linux filesystems benchmark, took a spare hard disk I had available and tested all the filesystems writable by my Linux system, a 2.6.33-gentoo-r2 kernel.

For the tests I used bonnie++ as root, with the command line _bonnie++ -d /mnt/iozone/test/ -s 2000 -n 500 -m hpfs -u root -f -b -r 999_.

My Linux system suppots a quite big number of writable filesystems, being: AFFS, BTRFS, ext2, ext3, ext4, FAT16, FAT32, HFS, HFS+, HFSX, HPFS, JFS, NTFS, Reiser, UFS, UFS2, XFS and ZFS.

I will divide the filesystems in "foreign unsupported", "foreign supported" and "native" filesystems.

A "foreign unsupported" is a filesystem designed and created outside Linux, for another operating system, but it's lacking features needed by Linux to run flawlessly, like the POSIX permissions, links and so on. These are AFFS, FAT16, FAT32, HFS and HPFS.

A "foreign supported" is a filesystem designed and created outside Linux, for another operating system, but supporting all features needed by Linux to run flawlessly, even if the current Linux implementation do not implement them fully. These are HFS+, HFSX, JFS, NTFS, UFS, UFS2, XFS and ZFS.

A "native" is a filesystem designed by the Linux community of developers, with Linux in mind, and usually never found outside Linux systems. These are BTRFS, ext2, ext3, ext4 and Reiser.

AFFS, or Amiga Fast File System is the old AmigaOS 2.0 filesystem, still in use by AROS, AmigaOS and MorphOS operating systems. It does not support the features required by Linux and is not an option for disks containing anything but data.

BTRFS, or B-Tree File System is a filesystem specifically designed to substitute old Linux native filesystems, including online migration of data from the extX filesystems.

ext2, or Extended 2 File System is an evolution of the original Linux ext filesystem, designed to overcome the limits of the Minix filesystems used in first Linux versions. ext3 is an evolution that added journaling and ext4 is another evolution that changes some structures to newer and faster technologies.

FAT, or File Allocation Table, is the original Microsoft filesystem, introduced with Microsoft Disk BASIC in the 80s and extensively used for data interchange, mobile phones, photo cameras, so on. FAT16 is an evolution of the original FAT (retroactively called FAT12) to allow for bigger disks, and FAT32 is the ultimate evolution introduced in Windows 95 OSR2 to overcome all the limitations. However Microsoft abandoned it in favor of NTFS for hard disks and ExFAT (currently not writeable by Linux) for data interchange.

HFS, or Hierarchical File System, is the original Apple filesystem introduced with Mac OS 5.0. It's design is heavily influenced by Mac OS design itself, requiring support for resource forks, icon and window position, etc.

HFS+, or Hierarchical File System Plus is an evolution of HFS to modernize it including support for UNIX features, longer file names, multiple alternate data streams (not only the resource forks). Current Linux implementation of HFS+ does not support all features from it, like for example journaling. HFSX is a variant of HFS+ that adds case-sensitive behavior, so that files "foo" and "Foo" are different ones.

HPFS, or High Performance File System is a filesystem designed by Microsoft for OS/2 to overcome the limitations of the FAT filesystem.

JFS, or Journaled File System is a filesystem designed by IBM for AIX, a UNIX variant, and so it supports all the features required by Linux. Really it's name is JFS2, as JFS is an older AIX filesystem that was deprecated before JFS2 was ported to Linux.

NTFS, or New Technology File System, is a filesystem designed by Microsoft for Windows NT to be the server-perfect filesystem. It was designed to support all features needed by all the operating systems it pretended to serve: OS/2, Mac OS, DOS, UNIX. Write support is not available directly in the Linux kernel but in userspace as NTFS-3G, and it does not implement all the features of the filesystem, neither all those required by Linux.

Reiser filesystem is a filesystem designed by Hans Reiser to overcome the performance problems of the ext filesystems and to add journaling. Its last version, Reiser4, is not supported by the Linux kernel because of reasons outside the scope of this benchmark.

UFS, or UNIX File System, also called BSD Fast File System is a filesystem introduced by the BSD 4 UNIX variant. UFS2 is a revision introduced by FreeBSD to add journaling and overcome some of the limitations. Support for both filesystem has been "experimental" in Linux for more than a decade.

XFS, or eXtended File System, is a filesystem introduced by Silicon Graphics for IRIX, another UNIX variant, to overcome the limitations of their previous filesystem, EFS. Linux's XFS is slightly different from IRIX's XFS, however both are interchangeable.

ZFS, or Zettabyte File System, is a filesystem introduced by Sun Microsystems for their Solaris UNIX variant. It introduces a lot of new filesystem design concepts and is still being polished. Support on Linux is provided by an userspace driver and not in the kernel itself.

The tests check the speed to write a 2 Gb file, rewrite it, read it, re-read it randomly, create 500 files, read them and delete them.

AFFS, FAT and HPFS filesystems failed the create/read/delete files test with some I/O error. In their native operating systems they have no problem to create 500 files so it's a problem in the Linux's implementation.

The output from Bonnie++ is:

Version 1.96		Sequential Output						Sequential Input				Random Seeks			Sequential Create						Random Create
	Size	Per Char		Block		Rewrite		Per Char		Block				Num Files	Create		Read		Delete		Create		Read		Delete
		K/sec	% CPU	K/sec	% CPU	K/sec	% CPU	K/sec	% CPU	K/sec	% CPU	/sec	% CPU		/sec	% CPU	/sec	% CPU	/sec	% CPU	/sec	% CPU	/sec	% CPU	/sec	% CPU
AFFS	2000M			6014	2	5065	3			14887	5	115.2	9
	Latency			9545ms		10553ms				16000ms		842ms		Latency
BTRFS	2000M			26126	11	14151	8			69588	5	35.0	262	500	14	72	55300	99	13	56	14	72	35273	99	8	47
	Latency			431ms		626ms				166ms		6071ms		Latency	6067ms		1712us		32642ms		6632ms		934us		7227ms
ext2	2000M			27592	4	13381	3			57779	3	424.1	5	500	84	63	401150	99	114	0	98	70	745	99	97	28
	Latency			326ms		526ms				23808us		279ms		Latency	522ms		214us		31756ms		483ms		8831us		878ms
ext3	2000M			27190	9	13522	3			58337	4	392.0	6	500	177	1	293176	99	129	0	177	1	392430	99	154	0
	Latency			754ms		626ms				49656us		193ms		Latency	859ms		876us		1200ms		844ms		205us		1497ms
ext4	2000M			28852	5	13832	3			57969	3	366.6	5	500	246	2	281186	99	155	1	246	2	372310	99	202	1
	Latency			4026ms		658ms				97049us		183ms		Latency	903ms		869us		1061ms		900ms		183us		951ms
FAT16	2000M			28201	6	13317	4			55201	3	354.9	6
	Latency			326ms		230ms				103ms		269ms		Latency
FAT32	2000M			26065	6	13317	4			54575	3	333.0	6
	Latency			326ms		230ms				103ms		269ms		Latency
HFS	2000M			28761	5	13272	5			50507	11	324.8	3
	Latency			426ms		526ms				16808us		150ms		Latency
HFS+	2000M			28705	3	13201	2			57350	4	472.7	2	500	3864	29	3513	99	5342	96	5338	43	248171	99	7754	67
	Latency			426ms		426ms				18678us		281ms		Latency	100ms		341ms		3064us		16396us		165us		753us
HFSX	2000M			28805	3	13353	2			57481	4	482.1	2	500	3899	28	3547	99	5451	96	5431	43	255570	99	7859	66
	Latency			227ms		410ms				24543us		212ms		Latency	152ms		338ms		2542us		14442us		211us		1452us
HPFS	2000M			25198	19	13488	6			47790	11	159.1	4	500	6980	97	216652	99	15720	46	7393	82	329146	99	16662	56
	Latency													Latency	14546us		1853us		3224us		9178us		180us		1510us
JFS	2000M			28486	4	13855	2			59308	3	546.6	3	500	7836	25	379769	99	2289	8	218	2	392966	99	96	0
	Latency			679ms		626ms				34570us		136ms		Latency	179ms		316us		747ms		2523ms		558us		67889ms
NTFS-3G	2000M			28324	6	12480	3			55439	5	156.5	1	500	5822	12	11353	18	8403	14	6280	14	13918	18	1025	2
	Latency			52838us		127ms				20516us		296ms		Latency	130ms		144ms		444ms		108ms		25397us		427ms
Reiser 3.5	2000M			27557	11	13567	4			58166	5	462.2	9	500	1704	10	328480	99	1785	10	1498	9	398475	99	448	3
	Latency			626ms		683ms				14264us		164ms		Latency	1299ms		657us		1882ms		1269ms		207us		2843ms
Reiser 3.6	2000M			27478	11	13472	4			58109	5	459.7	9	500	1458	8	319832	99	1404	8	1284	7	395686	99	396	3
	Latency			626ms		526ms				13669us		202ms		Latency	2405ms		672us		2078ms		1424ms		205us		3293ms
UFS	2000M			4484	4	14018	4			57305	7	417.1	8	500	102	86	417263	99	1252	6	103	87	414744	99	176	57
	Latency			410ms		567ms				81779us		185ms		Latency	488ms		164us		110ms		469ms		537us		497ms
UFS2	2000M			4472	4	13333	3			56329	6	431.9	7	500	100	87	419216	99	1257	6	102	87	418432	99	161	53
	Latency			426ms		331ms				100ms		512ms		Latency	487ms		185us		111ms		467ms		51us		537ms
XFS	2000M			28707	4	13759	3			55647	4	408.4	5	500	559	5	70286	98	557	4	637	6	110524	99	102	0
	Latency			5226ms		530ms				29344us		260ms		Latency	1555ms		14286us		1897ms		1421ms		277us		8480ms
ZFS	2000M			19540	3	12844	2			32664	2	152.5	1	500	753	4	37453	10	142	0	781	4	399620	99	51	0
	Latency			2562ms		2926ms				261ms		2268ms		Latency	2531ms		61859us		2965ms		2800ms		203us		5209ms

The table is quite long so comparing graphs following:

This is the table for a single file sequential writing speed. The UFS and AFFS denote an abnormally slow speed, indicating a big need for optimization. The fastest ones are, without a big difference, ext4, HFSX and XFS.

This is the table for a single file sequential reading speed. AFFS needs optimization also here. The fastest one with a big difference is BTRFS and the second fastest one is JFS.

This two tables indicate the speed on writing and reading big files. If this is your intended usage, the recommended ones are BTRFS, ext4, HFSX, HFS+, XFS and JFS. However if you need the maximum portability between systems, HFS+ is the choice because it's readable and writable by more operating systems than the others.

The following graphs show a more normal filesystem usage, where a lot of files are created, read and deleted at the same time.

This test shows surprising results. Reiser takes the lead on native filesystems, while even the newest BTRFS is quite behind all the other native ones. In foreign filesystems, HPFS takes the lead. However, because of HPFS design limitations, the choice is NTFS, even being run on userspace.

In randomly reading 500 files, the results invert. NTFS loses its lead to UFS2. ext3 shows a huge improvement to ext2, and inversely, BTRFS shows a huge lose of performance.

Finally deleting all files, the lead is taken again by HPFS, second by HFS+ and HFSX.

Unless you want to write files and never delete them, the obvious choice is HFS+ or HFSX.

Surprisingly, even not being on its native implementation, HFS+ provides the best combination in all tests. It supports all of the features required by Linux, gives a good sequential speed, and maintains its speed when handling a big number of files.

The only bad thing is that Linux does not support the journaling. Not a problem if you take care on unmounting it cleanly, without "oops! what happen to the electricity".  Of course you can always spam the Linux kernel developers so they implement that function of the HFS+ filesystem that exists from 2003, 8 years ago.