CentOS virt - Aug 2012 - tuned-adm fixed Windows VM disk write performance on CentOS 6

On a 32bit Windows 2008 Server guest VM on a CentOS 5 host, iometer 
reported a disk write speed of 37MB/s

The same VM on a CentOS 6 host reported 0.3MB/s.  i.e. The VM was unusable.

Write performance in a CentOS 6 VM was also much worse, but it was usable.
(See http://lists.centos.org/pipermail/centos-virt/2012-August/002961.html)

With iometer still running in the guest, I installed tuned on the CentOS 
6 host and enabled the virtual-host profile.  iometer showed the average 
disk write speed increasing immediately. A fresh iometer test reported a 
disk write speed of 80MB/s.

I'm not sure if the tuned-adm virtual-host profile was available in 6.0 
- it may be necessary to update to get it.
http://docs.redhat.com/docs/en-US/Red_Hat_Enterprise_Linux/6/html/Power_Management_Guide/tuned-adm.html

Installation is quick & easy:
yum install tuned
tuned-adm profile virtual-host

I'm surprised not to see any other posts on this subject.  Maybe there's
something peculiar about my setup that caused the dreadful write 
performance & tuned-adm happened to fix it.  This problem has been a 
real headache and I'm extremely grateful to Philip Durbin and jimi_c for 
bringing the solution to my attention.

The host server was rebuilt from scratch after a RAID1 simultaneous 
double disk failure (happily my DR strategy worked. Have you tested 
yours recently?).  It appears that CentOS 6 has some different defaults 
(more conservative power saving?).  Maybe upgrades don't change the 
defaults, so this issue only affects new CentOS 6 installs?

The iometer settings were: 1000 sectors, 16 outstanding IOs, access 
specification: 16k, 0%read 0%random
The figures I've reported above are the 'Total MBs per second'
See http://www.iometer.org/doc/downloads.html

In the process of researching the problem, 2 other quick & easy 
optimisations came to light, each adding a few more MB/s:
- Set noatime. See: 
http://www.activoinc.com/blog/2009/08/25/setting-noatime-and-nodiratime-for-improved-disk-performance/

- Disable Windows disk cacheing: See 
http://pve.proxmox.com/wiki/Windows_2003_guest_best_practices
On Windows 2008: Computer Mgmt -> Disk Mgmt -> Select Drive (system) ->
properties -> Policies tab -> uncheck 'Optimize for quick removal'

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.centos.org/pipermail/centos-virt/attachments/20120812/f54a9fb4/attachment-0006.html>

> I'm surprised not to see any other posts on this subject.  Maybe
there's
> something peculiar about my setup that caused the dreadful write 
> performance & tuned-adm happened to fix it.  This problem has been a 
> real headache and I'm extremely grateful to Philip Durbin and jimi_c
for
> bringing the solution to my attention.
     I don't know about 6.0 because I went from 5.something to 6.2, then 6.3
and tuned was installed by default but not with those profiles.  With CentOS
6.2, I manually tuned things, played with read-ahead sizes, etc, until I got
them optimized.  Your milage may vary but with a machine with 32GB of RAM, I
got the best performace with a read ahead of 16384 blocks.  I set the I/O
schedualar to deadline to limit latency.  I played with various cache modes and
for most of my applications write-back was okay (i.e., not storing critical
financial data and crashes are few and far between, journalling file systems
are used, so worst case someone might lose a couple minutes work).

     But that said, one weirdness I have run into while playing with options to
get latency to a minimum, I found that if I enabled kernel premption, the
kernel would only use one core.  Is there anyway to have a preemptive kernel on
a multicore machine and still use all the cores?