Phoronix: Benchmarks Of The New ZFS On Linux: EXT4 Wins
At the end of March was a new release of ZFS On Linux, a kernel module implementation of the ZFS file-system for Linux, and it was declared ZFS On Linux is now ready for wide-scale deployments. With this release (ZOL/SPL v0.6.1), new benchmarks are being done to compare ZFS to popular Linux file-systems. In this article is a brief preview against EXT4...
Well, ryao won't be pleased with this article in the slightest.
Isn't this like comparing an armoured truck versus a sportscar with central locking?
From the ZFS on Linux FAQ:Additionally, it should be made clear that the ZFS on Linux implementation has not yet been optimized for performance. As the project matures we can expect performance to improve.
then it's at least partly comparable
when backing up my 1.2-1.5 TB of data from an ext4 home-partition to ext4 vs. btrfs & ZFS
btrfs (with gzip-compression) takes about 2x the time than ext4 and approx. 1.5x the time of XFS (strangely xfs got slower with my latency-related tuning & ext4 got faster )
ZFS with lz4 and/or lzjb compression takes about 2.5x-3x the time of ext4
and that's only the raw transfer-time (8 minutes of ext4, approx. 30 minutes of ZFS) and neglecting all the data-safety wise features of e.g. ZFS:
- sha256 checksums (said to be the slowest compared the default hash algorithm, lower probability of hash collisions with e.g. deduplication)
- checksums on the whole filesystem - from file, metadata up to higher parts of the filesystem
- optional single-filesystem dedup & multi-copy functionality (didn't know about that until recently and will use it on critical study/personal files): e.g. copies=2
- there's several more - and I'm not really an expert in filesystems so just let your google-fu take you to wikipedia, etc. to find out more
looking forward to it
The tests were done with early Intel SSDs that are known to lie about their block size, presumeably without making any adjustment to ashift. ext4's assumption that everything is 4KB is advantageous here. In additional, 3 of the 4 tests appear to be read/write once tests, which do not reflect real world workloads. I say 3 of 4 because I have not read enough about dbench to say one way or another, although I can say that the single client configuration is likely the worst possible test possible.
Last edited by ryao; 04-18-2013 at 06:05 PM.
I find it a little bit useless to just compare ZFS against EXT4.
EXT4 is interesting as a point of comparison. But they don't even belong to the same generation.
EXT4 is basically EXT3 with extents added in (which in turn was EXT2 with journalling thrown in).
It an previous-gen filesystem.
Where as ZFS is a newer generation file system with a ton of features (with checksumming of nearly everything, data de-duplication, redundancy, on-the-fly compression, copy-on-write, subvolumes and snapshots, and a whole volume management).
It should be compared against similarly full-of-feature files system (btrfs more or less covers the same feature set).
ZFS, BTRFS and Co vs. EXT4 and Co are never going to run at the same speed, given that the first have much more features that require some CPu cycles. And they don't target the same audience.
The current benchmark would be like benchmarking FAT32 vs. a distributed cluster file system. Even if they are all filesystems, they don't have the same feature set at all and don't aim for the same users.