Linux 3.11 File-System Performance: EXT4, Btrfs, XFS, F2FS

[phoronix]

Phoronix: Linux 3.11 File-System Performance: EXT4, Btrfs, XFS, F2FS

Coming out today are our first Linux 3.11 kernel file-system benchmarks. Being benchmarked from a higher-end OCZ Vertex 3 SATA 3.0 SSD connected to an Intel Core i7 "Haswell" system are the EXT4, Btrfs, XFS, and F2FS file-systems.

http://www.phoronix.com/vr.php?view=18936

[jwilliams]

Since many of your disk benchmarks are writing zero-fill data by default, you should NOT be using a Sandforce SSD to do these sorts of tests, since Sandforce compression benefits unrealistically when given zero-fill data instead of realistic data.

[sheldonl]

Include LVM (Logical Volume Manager) in the FS benchmarks.

I've seen a lot of phoronix fs benchmarks, and what I'd like to see as a fair comparison to the btrfs benchmarks is ext4+LVM and xfs+LVM. It always appears that brtfs is so much slower than the other file systems, so users go and use ext4 or xfs or something else but then also use LVM so that they have the flexibility of managing their storage. However, I've seen the use of LVM incur up to a 30% hit on performance. Being that btrfs includes this type of functionality I think a much fairer comparison would be to include LVM when benchmarking the other file-systems.

Michael, what do you say?

[halo9en]

ext4 is still the most performant, but why is it getting worse (fs-mark, dbench, compile bench) with every kernel release?

[blackout23]

Since many of your disk benchmarks are writing zero-fill data by default, you should NOT be using a Sandforce SSD to do these sorts of tests, since Sandforce compression benefits unrealistically when given zero-fill data instead of realistic data.

Not only that, but he'll better buy a second SSD anyway. OCZs SSD with Sandforce are dying like flies, because they are so shitty. Happend to the KWin developer just a few days ago.

https://plus.google.com/u/0/115606635748721265446/posts/CcdV1R67bpe

[FourDMusic]

ext4 is still the most performant, but why is it getting worse (fs-mark, dbench, compile bench) with every kernel release?

Since this is seen across the board for the different file systems, I'm wondering if there is a regression in the VFS layer or something else that lives underneath the FS layer. Could an FS expert comment on this?