freebsd stable - Nov 2015 - ZFS - poor performance with "large" directories

On Tue, Nov 24, 2015 at 8:00 AM, Albert Cervin <albert at acervin.com>
wrote:
> Hi all,
>
> Please feel free to direct me to a list that is more suitable.
>
> We are trying to set up a fileserver solution for a web application that we
> are building. This fileserver is running FreeBSD 10.2 and ZFS. Files are
> written over CIFS with Samba running on the fileserver host.
>
> However, we are seeing en exponential decrease in performance to write to
> the file server when the number of files in the directory grows (when it
> goes up to ~6000 files it becomes unusable and the write time has gone from
> a fraction of a second to ten seconds).
>
> We ran the same setup on a Linux machine with an ext4 file system which did
> NOT suffer from this performance degradation.
>
I should hope not.  ext4 vs zfs comparison isn't fair for either.

>
> Are these "holes" in write speed normal. Since this is the exact
symptom we
> are getting when the network writes start to be slow.
>
Totally normal.  You'll want to reference:

https://wiki.freebsd.org/ZFSTuningGuide

In particular for that issue see:

vfs.zfs.txg.timeout
and tuning related to NFS.

Performance is also heavily dependent on pool structure and io
characteristics.  For example, a pool of 3 2 disk mirrors is in general
going to be much faster than 1 6 disk raidz2.


-- 
Adam

Thanks!

"I should hope not.  ext4 vs zfs comparison isn't fair for
either."

I do realize that comparing ext4 and ZFS is not really giving anything
but it tells us one thing, ext4 would work whereas ZFS would not for
our use case, which was unexpected, at least to me.

 vfs.zfs.txg.timeout is already verified to be 5 (the default). I have
also turned off atime and vfs.zfs.arc_meta_limit is 1287906304.

"Do you have any memory pressures on your server ?  Have a look at this
thread"

The server has 4 cores and 8Gb of memory and there is no memory
pressure on the server. Thanks, will look at the thread!

The setup is _really_ simple with just a single pool and no datasets.

Any more ideas?

Cheers,

Albert