About :: TODO :: Blog :: RSS :: Old blog :: Projects :: GIT :: Gallery :: Notes

Fri, 27 Apr 2007

Great filesystem contest. Introduction.


What is your favorite one?
So far, only postmark tests completed for default mounted filesystems, and also couple of dbench/iozone for default and tuned mounts, and even those results are interesting in some regard.
Let's speculate a bit about postmark benchmark.
This one has been completed for all filesystems for default mount options so far.
This benchmark performs set of random read/write/create/delete operations, each file is usually quite small (size can be random (as in test) or obey some rules for maximum and minimum size (where postmark has some bug which leads to crash, so I did not use that option)). This benchmark emulates SMTP/NNTP/small-e-commerce workload, where system operates with set of relatively small files.
Of course we know that xfs and jfs are bad choise for such load, and the best choise is likely reiserfs (I did not yet tested Reiser4, but I want to).
But... when amount of files is relatively small (10k files) reiser is not that good as we expect, with 30k files it is a winner, but with small difference (maximum 10% though).
We also expect ext4 to be better or at least not worse. This is not entirely correct too. Looking into graphs I would not say I want to use ext4 instead of ext3.

I removed ext2 results (will lilely redo them), since they are too good to be true - in order of 3-10 times better than the nearest competitor, sometimes its speed is in order of 2 times higher than disk speed, which means excessive usage of the page cache, which never shows in ext3/4, so I dropped - will redo later.

So, stay tuned, I will complete other tests (including tuned mount options, mostly notail, writeback journalling and extents) and present graphs here.

Tests are not aimed to show which filesystem is the best for which workload, but to find existing performance limits as a target.

/devel/fs :: Link / Comments ()