[CentOS] Disks for virtualization

Fri Jul 16 02:47:26 UTC 2010
R P Herrold <herrold at owlriver.com>

On Thu, 15 Jul 2010, David Dyer-Bennet wrote:

> Much the easiest to manage is just having the virtual machine use a file
> in dom0 as its disk.  Without pre-allocating, this lets me over-commit
> somewhat to cover unknown future needs, for example.

One sees this a lot in the writeups; one assumes that 'zoned' 
updates within the filesystem image are being done, rather 
than the whole image being repeatedly re-written

> The alternative seems to be LVM partitions.  Those are easy to throw
> around too, but after enough creation / destruction you can end up with
> your free-space fragmented so you can't use it, right?

I don't know that one can see it -- one reason for the logical 
volume abstraction is to provide information hideing of 
free-space fragmentation.  I assume you read this somewhere, 
but we have not encountered with the PMman [1] VM's as an 
issue we have needed to address

> But I imagine there's a performance benefit to LVM partitions over dom0
> files.

We concluded that using LVM was more complex to write the 
front end to as to deployment matters, but that it gave us 
native filesystems performance (mod the LV layer), and that we 
could manipulate those filesystems when quiescent 'out of 
band' to the hosting domain zero.  We have ended up writing 
lots of local tools, but then we know we would be writing code 
anyway to support our UI

I know we seem markedly faster than AWS, Rackspace and prgmr 
as to deployments and backups, but have no metrics to show 
this formally.  Certainly wayy faster than the s390x unit I 
work on (a VM under a hypervisor of a wholly different class, 
and designed for different load profiles and scaling)

> For things like foswiki for internal use by a development team, and Nagios
> monitoring for about a dozen systems, do i need that extra performance?

probably not; We run sysstat (sar) recording on all production 
units, and we rarely budge the load factors materially, with 
excellent latency profiles.  I would prototype it first; 
contact me out of band and let's discuss details of you doing 
that in a donated instance, pre-production, and documenting 
your results back to us here
 	'premature optimization is the root of ... '
 		-- D. Knuth

> Does anybody have a rule-of-thumb for the difference?

I looked and you had drawn no responses -- just my opinions ;)

-- Russ herrold
 	also: herrold at pmman.com
[1] http://www.pmman.com/