ext3 performance -- very large number of files, large filesystems, etc.

td3201 · 11-24-2008, 10:10 PM

Hello all,

I have a decent sized NFS cluster (18 TB) on RHEL5. It is spread out with 1-4 TB volumes of ext3 on top of LVM. I am seeing questionable performance. It is a backup application that uses the disks so it is 75% write. Each file system has a lot of files (>15,000,000). All that being said, I am curious how to further tune this system. The volumes were created and mounted with default options.

My first plan is to mount them with data=writeback. Any other ideas?

thedonkdonk · 11-25-2008, 12:31 AM

you might try reducing the inode size.

PEdroArthur_JEdi · 11-25-2008, 08:40 AM

Quote:

Originally Posted by td3201

Any other ideas?

http://linuxgazette.net/122/TWDT.html#piszcz

See the above link about a benchmark on filesystems. Maybe your answer is there!

td3201 · 11-25-2008, 08:47 AM

Ya, that shows that ext3 performance sucks when created out of the box with no modifications. I've established that.

td3201 · 11-25-2008, 09:09 AM

Quote:

Originally Posted by thedonkdonk

you might try reducing the inode size.

It's currently set to 128 (default). There is no way to reduce the inode size on a live volume, no?

PEdroArthur_JEdi · 11-25-2008, 09:28 AM

Quote:

Originally Posted by td3201

Ya, that shows that ext3 performance sucks when created out of the box with no modifications. I've established that.

I think the most important in the link I point you is try to figure out which is the most frequent operation in your server. If files are created all the time, then ext3 will suck. It's file permission and mactime mechanism is too overhead. You may read an official discussion about this topic at http://kerneltrap.org/node/14148 . But, if the files are just being updated (like if you have a large set of already created files), mounting the filesystem as relatime will handle ok.