On Mon, Jan 23, 2012 at 01:19:29PM -0500, Josef Bacik wrote:

On Fri, Jan 20, 2012 at 01:13:37PM +0100, Christian Brunner wrote:

quoted

As you might know, I have been seeing btrfs slowdowns in our ceph
cluster for quite some time. Even with the latest btrfs code for 3.3
I'm still seeing these problems. To make things reproducible, I've now
written a small test, that imitates ceph's behavior:

On a freshly created btrfs filesystem (2 TB size, mounted with
"noatime,nodiratime,compress=lzo,space_cache,inode_cache") I'm opening
100 files. After that I'm doing random writes on these files with a
sync_file_range after each write (each write has a size of 100 bytes)
and ioctl(BTRFS_IOC_SYNC) after every 100 writes.

After approximately 20 minutes, write activity suddenly increases
fourfold and the average request size decreases (see chart in the
attachment).

You can find IOstat output here: http://pastebin.com/Smbfg1aG

I hope that you are able to trace down the problem with the test
program in the attachment.

 
Ran it, saw the problem, tried the dangerdonteveruse branch in Chris's tree and
formatted the fs with 64k node and leaf sizes and the problem appeared to go
away.  So surprise surprise fragmentation is biting us in the ass.  If you can
try running that branch with 64k node and leaf sizes with your ceph cluster and
see how that works out.  Course you should only do that if you dont mind if you
lose everything :).  Thanks,

Please keep in mind this branch is only out there for development, and
it really might have huge flaws.  scrub doesn't work with it correctly
right now, and the IO error recovery code is probably broken too.

Long term though, I think the bigger block sizes are going to make a
huge difference in these workloads.

If you use the very dangerous code:

mkfs.btrfs -l 64k -n 64k /dev/xxx

(-l is leaf size, -n is node size).

64K is the max right now, 32K may help just as much at a lower CPU cost.

-chris