[Gllug] performance of xen dom0 vs native linux

[Richard Jones]

On Sat, May 09, 2009 at 11:50:28AM +0100, James Courtier-Dutton wrote:
> Some of them use more so called "hardware acceleration" such as Intel
> VT or AMD-V.
> It has been shown that using hardware acceleration can actually slow
> some applications down when compared with software emulation,
> particularly with the intel vt and amd-v hardware instructions.
> Example applications are ones that use a lot of malloc/free calls.
> intel-vt and amd-v can slow these applications down quite a lot.

James, extraordinary claims such as this one require extraordinary
proof, which I'm sure you'll be happy to provide.  I'm interested to
know how exactly userspace library details like malloc/free are
supposed to influence guest kernel performance.

I will also note here that there are different levels of hardware
assistance available, and Intel's and AMD's approaches also have
different performance characteristics.  This webpage for instance
explains Intel's earliest approach (pre-EPT):

http://download.intel.com/technology/itj/2006/v10i3/v10-i3-art01.pdf

[...Digression into dosemu clipped...]

> Para-virtualisation uses a good approach by instead of having to
> emulate actual devices, it provides a special device driver for the
> guest that improves performance when talking to devices. Emulating
> actual devices leads to performance issues due to the double buffering
> etc. required. Para-virtualisation can help considerably for
> applications that use a lot of device io.

Paravirtualization is more than just devices.  This paper explains
Xen's approach to paravirtualization RAM and devices:

http://www.cl.cam.ac.uk/research/srg/netos/papers/2003-xensosp.pdf

Rich.

-- 
Richard Jones
Red Hat
-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug