Journalling File System Comparison

Ithika writes "Justin Piszcz recently did an analysis of some common journaling file systems over at Linux Gazette. Due to the Gazette's ridiculous restrictions on image filesize all of the graphs are pretty much illegible, however. You can see the article as the author originally intended here."

Comments

Possible Explanation for Bad Dir Numbers?

[man1ed]

Regarding the bad performance of ext3 on directories, is this feasible?:
Since every directory has two entries in it ("." and ".."), that is three times the number of directory entry creations/deletions that need to be journaled. Perhaps the other journaling FSs have some optimization for handling "." and "..".

Use the filesystem that fits

[Profane MuthaFucka]

It's important to pick a filesystem that performs the main task asked of a filesystem. Keeping that criteria in mind, it's easier to make sense of these results.

Pick the filesystem that won't trash your data. I haven't seen any benchmarks that show stats like errors in the filesystem code, maturity, or how often the filesystem corrupts unexpectedly.

ext3

[Anonymous Coward]

I must say that ext3 has had major improvements in 2.6 (no, I know that won't help people running 2.4 so you don't need to tell us again).

I'd still use it for any system unless I could demonstrate that it really couldn't handle the workload.

Then again with the hard disk being a bottle neck quite often in desktop systems, there is something to be said about a fast filesystem there.

Problem with ext3

[kompiluj]

In fact you have not one but three ext3's: data=journal, data=writeback, data=ordered (look here for details) with different performance.
So if anyone tries filesystems comparison, please benchmark all three data modes of ext3 or, at least, provide information which one was used in the test.
To add my 0.02 EUR I would also like to see comparison of linux filesystems against *BSD ufs2 on the same machine. I know that the underlying operating system adds additional variables in the equation, but if you are interested in filesystem-oriented tasks the system's performance as a whole counts, not only its filesystem. I have seen something like this done already for RAIDs but I would like to see this for ordinary single drives (like ATA and SATA). Shall I have enough time I will perform such tests end of July, when my new hardware arrives.