6-Way File-System Comparison On The Linux 4.1 Kernel

[speculatrix]

I'm surprised that XFS doesn't really seem to excel at anything.

[rdnetto]

The really interesting comparison here is btrfs vs NILFS2, because those are the only two benchmarked which provide snapshotting facilities. (When actually used for snapshotting, I suspect btrfs would have an even greater performance hit.)

FWIW, I use btrfs on all my systems, despite being slower, because being able to rollback arbitrary changes is just so damn useful. (I also have a XFS partition on an SSD for anything IO intensive.)

[mcdebugger]

I'm using ext4 in almost all setups but I'm getting acustomed to the btrfs. Maybe in future I'll switch to it but for now ext4 is enough good for me.
As of some fses' special features I'm happy with mdraid + LVM + ext4 construction and don't plan to switch to some other setups now.

[Sloth]

For databases? In the case of automated tests, etc, or if that partition is just used for database storage only.

You can mount your entire partition with data COW enabled, but only disable COW for database directories with chattr +C (and then copying the relevant files, since chattr +C only effects newly created files.)

I'd be very interested in seeing the postgres portion of these tests, for example, run with just the postgres data directory setup with data COW disabled. It would be very interesting to compare that to the current results.

[stiiixy]

For this particular test, no, as I only had time to test this particular SSD on Linux 4.1 before re-tasking it to a different system, etc. It really depends on the interest level and related interest of recent articles whether it's a multi-way kernel comparison article, etc.

I think you will find a lot of people genuinely like reading up on filesystem benchmarks; and not just for performance results. As was noted, a lot are using EXT4, XFS and BTRFS (or wanting to), all for differing reasons. I myself am curious how shitty old FS's like FAT32 compare to F2FS, as I posted previously in another thread, simply for pure raw speed for file loads but also how stable some systems are under certain loads. Probably jsut easier to poke through the openbenchmarking results, ay =)

I use EXT4 for system partitions (and friend PC's), XFS for large media files, and whatever OpenELEC uses (could care less, as it's all network storage anyway).

Thanks for the article.

[user82]

Would be great to see another test of compression methods with btrfs on a reasonable notebook like machine.
Are there cases where the cpu bottlenecks, maybe in some the ssd is the bottleneck.

[Michael]

Thank you Michael, I really like these articles. Would it be possible to include in the same graphs for each file system 2 bars, one for the previous linux version to compare?

For this particular test, no, as I only had time to test this particular SSD on Linux 4.1 before re-tasking it to a different system, etc. It really depends on the interest level and related interest of recent articles whether it's a multi-way kernel comparison article, etc.

[Michael]

Why mount btrfs nodatacow instead of using chattr +C on the relevant directory?

For databases? In the case of automated tests, etc, or if that partition is just used for database storage only.

[s_j_newbury]

Why mount btrfs nodatacow instead of using chattr +C on the relevant directory?

[bug77]

I was using xfs, but recently switched to ext4, with plans for switch to btrfs(I will wait for it to stabilize a bit before the jump)

I was happy with xfs, too, until I messed up my partition table (my fault, not file system's). At that point I have discovered that while there are tools that will help you recover an ext partition, hardly any will work for xfs. Plus, as I read in the meantime, xfs's strength is handling huge files. And I don't edit video.

Back to the article, it's nice, but the picture is more complex. I'd throw a mechanical drive in there, because the best file system for a SSD will not necessarily perform the same on a HDD. And many people still use those for storage. I'd also test using SSDs with controllers from different manufacturers, there may be a few surprises in that area.