[Gluster-users] GlusterFS for production use with Ovirt

[Peter B.]

On 03/24/2015 04:01 AM, Paul Robert Marino wrote:
> 5) Look closely at the scheduler on the disks the default these days
> is CFQ (Completely Fair Queuing). I find DEADLINE works better for
> Gluster; however depending on if you are using a RAID and you have a
> properly aligned file system you may find NOOP may provide better
> performance you need to do some testing with your hardware to
> determine if this is the case.

We are running at setup here at the AV archive where glusterd is running
on RHEL 6.6 as KVM guest.
I've done some performance tests, and the scheduler settings we've found
to be the fastest in our setup are:

vm-host (Debian):
  kernel: elevator=deadline

vm-client (RHEL):
  kernel: elevator=noop
  qemu/libvirt: cache=none, io=native


The bricks consist of 15 SATA disks each using Linux kernel software
RAID-6 (1 spare) on the KVM-host (not in the VM), formatted with XFS.
More details about my performance tests can be found here:
http://www.das-werkstatt.com/forum/werkstatt/viewtopic.php?f=7&t=2109


(NOTE: The actual MB/s speeds are not that amazing, but that's due to
hardware limitations)


Regards,
Pb