I'm currently using the traditional rsync+cp -al method to create incremental/snapshot backups of our server tree. The backups are going onto a pair of eight-disk towers connected to the backup machine (a Sandy Bridge machine with 16 GB of RAM, running CentOS 5.5) via four eSATA connections (four disks per connection). Each disk is a regular 2 TB disk, so we have 32 TB of disk space connected to the backup machine. We're backing up about 20 TB of data on the servers with this.
The problem is that each daily backup is taking more than 24 hours, and the real time-killer isn't the actual rsync, but the time it takes to perform a cp -al of the tree locally on the backup machine. It's taking more than 12 hours just to make the shadow copy of the tree, and as far as I can tell the performance backlog is at the disk (top shows the cp using a lot of RAM but not a lot of CPU and mostly in uninterruptible-sleep state)
We have the server data split into four major volumes (and a few minor ones), and each of these backups runs in parallel (with some offsets in the cron to try to get some disks' cp done first). There are two volumes on the backup drive, both striped LVM volumes of 16 TB each.
So obviously I need to improve the performance because it's unusable as it stands.
The first question is: when CentOS 6 comes out, with support for btrfs, will making snapshots of subvolumes with btrfs substantially increase this performance?
The second is: is there a way, with ext3 or something else supported in CentOS 5 or 6, to 'encourage' it to put the directories/inodes in one part of a volume (which could happen to be the part that's on an SSD, via LVM) and the files in another? That would presumably solve the problem, but I don't know of ways to hint ext3 like that.