[Discuss] Using raw host hard disk in virtual client

[richard.pieri at gmail.com (Rich Pieri)]

On Sat, 4 Aug 2012 18:22:34 -0700
"Rich Braun" <richb at pioneer.ci.net> wrote:
> Performance is equal to or possibly better than the bare-metal,

This is not possible.  Even with raw disk I/O the guest does not talk
directly to the disk controllers.  It talks to the emulated controllers
exposed by the host.  This emulation incurs a small processing overhead
so virtualized I/O can never be as fast or faster than bare metal.

That said, the guest also has a disk I/O cache of its own.  Certain
kinds of disk I/O will remain within the guest's cache.  This could be
the cause of your perceived faster than bare metal performance.  Try
mounting the guest file systems sync and see what happens.

> setting it up this way.  (How could it be "better"?  The big host OS
> cache, which should be turned off if you're concerned about data
> integrity during a power outage.)

I've tried a configuration like this.  Debian 6 64-bit host with
VirtualBox from Squeeze backports.  ext3 file system for the guest
containers mounted sync specifically to bypass host I/O caching.  Guest
containers pre-allocated to full size.  I experienced two severe
problems.  The first is that this configuration was noticeably slower
than with the host file system mounted async.  The second is that the
guest VMs kept crashing under load with the file system mounted sync.
This crashing problem disappeared when I reverted to async mounts.

Seriously?  If you're concerned about I/O performance and file
durability then you shouldn't be running critical storage inside
user-mode hypervisors.  Move the storage to bare metal even if you have
to run it directly on the host.  You'll be better off for it.

-- 
Rich P.