[CentOS] is Intel VT-d "really" necessary?

Thu Sep 16 19:12:52 UTC 2010
Pasi Kärkkäinen <pasik at iki.fi>

On Wed, Sep 15, 2010 at 11:19:38AM +0200, Rudi Ahlers wrote:
> Hi all,
> 
> I'm just curios and would like some input from the community on this
> one. We're busy budgeting for a couple of new servers and I thought it
> would be good to try out the Core i7 CPU's, but see the majority of
> them don't offer VT-d, but just VT-x. Looking at the LGA1366 range,
> only the "Intel lga1366 i7 980XE" (from the list of what our suppliers
> stock) have VT-d, and it costs 4x more than "Intel lga1366 i7 930" or
> 2x more than "Intel lga1366 i7 960". From a budget perspecitve I could
> purchase 4 more CPU's, which could translate to 40x - 80x more VM's
> being hosted for the same capital outlay. Experience has shown that we
> under-utilize CPU's by a great margin and memory / HDD IO is our
> biggest bottleneck on any server.
> 
> So, if VT-d really necessary?
> We mainly host XEN virtual machine for the hosting industry, i.e. we
> don't need / use graphics rendering inside VM's, or need DAS on the
> VM's, etc.
> 

VT-d is marketing term for Intel's IOMMU (IO MMU) implementation,
and it's used *only* for PCI passthru, aka giving guest VM direct PCI
access to some physical PCI device (nic, hba, etc) on the host hardware.

Xen can actually do PCI passthru *without* VT-d for PV guests,
but for Xen HVM guests you *need* VT-d (if you want to use PCI passthru).

VT-d is NOT required for running HVM/Windows guests.

VT-x is the CPU feature that makes it possible to run unmodified guests.
VT-d is the chipset IOMMU feature for PCI passthru.

See:
http://wiki.xensource.com/xenwiki/XenPCIpassthrough

-- Pasi