I'd like to see more filesystem choice

Story: Ext3 for large filesystems and Time for ext4?Total Replies: 2
Author Content

Jun 24, 2006
12:15 PM EDT
I recently benchmarked ext3, reiser3, jfs and xfs on a new server, and almost as an afterthought, I tested ext2 as well. I ran bonnie, dbench, tiobench and iozone on each file system.

The results were interesting, for the most part backing up the sorts of things I had heard over the past few years of filesystem benchmarks.

Overall reiserfs seemed to consistently give the fastest results, while jfs was slowest over most of the range of tests, but seemed to hold up pretty well under heavy load.

jfs also used the least CPU, almost as if it's specifically being throttled.

xfs showed the lowest latency, and the greatest consistency of throughput.

ext3 was all over the map - at times it was fairly close to reiser in speed, at other points it was maybe 20% of the speed of reiser.

When I ran ext2, I was somewhat surprised to see how badly it whipped all the others in every single test.

For instance, ext2 had a pretty consistent throughput of 1000Mb/s over a wide range of dbench client settings, while the others ranged from 100 to around 400 MB/s.

Bottom line, I would love to see a robust modern filesystem for linux, with the excellent performance of the lean, mean ext2.

Jun 24, 2006
6:19 PM EDT
Well, sure ext2 is going to win on throughput, it doesnt do journaling. The way the other filesystems do journaling involves the metadata (and data if you have that option enabled) being written twice. ext2 is stll a perfectly valid choice, expecially if raw throughput is the main requirement. (and fscking a 100G full filesystem isnt going to ruin your day.) btw. I would be curious if your benchmarks of the ext2/3 filesystems involved the 'dir_index' feature being enabled? Could make a difference in some uses.

Jun 25, 2006
12:32 PM EDT
To compare apples to apples, you should mount the ext3 fs with "-o data=writeback". ext3 makes more integrity guarantees than XFS or JFS at the expense of speed. IIRC, reiser3 also makes those guarantees by default now. (I could be wrong on that, but I know it supports "ordered" mode which is the mode in question.)

Personally, though, for production use, I would keep ext3 in the default data=ordered mode unless speed were absolutely critical. To me, the reduced guarantees that JFS, and in particular XFS give, are too much like playing Russian roulette with my data.

Posting in this forum is limited to members of the group: [ForumMods, SITEADMINS, MEMBERS.]

Becoming a member of LXer is easy and free. Join Us!