Xen devel - Jan 2009 - paravirtualized vs HVM disk interference (85% vs 15%)

Hi,

We have found that exist a huge degradation in performance when doing I/O to a
disk images contained in single files from a paravirtualized domain and from an
HVM at the same time.

The problem was found in a Xen box with Fedora 8 x86_64 binaries installed (Xen
3.1.0 + dom0 Linux 2.6.21). The test hardware was a rack mounted server with two
2.66 Ghz Xeon X5355 (4 cores each one, 128 Kb L1 cache and 8 Mb L2 cache), 4Gb
of RAM and one disk of 250 Gb.  Both paravirt and HVM domains have 512 Mb of RAM
and 8 vcpu''s and runs also a Fedora 8 x86_64 distro.

Stressing at the same time the paravirtualized and HVM guests with disk
stressing tools like ''stress'', bonnie++ or hand written
dd''s, the amount of disk activity is not fair shared among both. The
paravirtualized domain shows on average 85% of I/O transfer rate in all the
tests, against a poor 15% for the HVM domain (the iotop tool was used in dom0 to
obtain the data about I/O transfers for qemu-dm and loopX processes).

Our 2 questions are:
- The same problem exists in the dom0 2.6.18 kernel? We cannot use the iotop
tool to measure this on 2.6.18 since I/O per process accounting is not supported
on it. However we suppose that could be possible put the two disk images in
different HDD''s and use the iostat tool from the sysstat suite (which
allow to sees disk activity in a per-device basis) to observe the behavior.
- Is the behavior ok? We mean, we consider that this interference is undesirable
for certain deployments (anyone with paravirtualized and HVM guests in the same
box, like in example Linux+Windows2003, and you want a fair share of the disk
transfer rate). What we are asking is if this is a consequence of the Xen design
and a known behavior, and if there is a workaround to ameliorate the
interference (the problem persists and in the same degree even if both guests
perform writes below dom0''s dirty_ratio).

Regards,

Duilio J. Protti,
Alejandro E. Paredes

Intel Argentina Software Development Center


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

On 27/01/2009 05:38, "Protti, Duilio J"
<duilio.j.protti@intel.com> wrote:
> - Is the behavior ok? We mean, we consider that this interference is
> undesirable for certain deployments (anyone with paravirtualized and HVM
> guests in the same box, like in example Linux+Windows2003, and you want a
fair
> share of the disk transfer rate). What we are asking is if this is a
> consequence of the Xen design and a known behavior, and if there is a
> workaround to ameliorate the interference (the problem persists and in the
> same degree even if both guests perform writes below dom0''s
dirty_ratio).
Run PV drivers in the HVM guest. This will automatically mean the HVM guest
is using a more efficient data channel, of the same sort as the PV guest. If
you need further fairness guarantees you can make use of one of Linux''s
fair
queuing disc schedulers.

 -- Keir



_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel