[Gllug] best way to update a single production server?

Karanbir Singh mail-lists at karan.org
Fri May 1 11:02:45 UTC 2009


John Hearns wrote:
>> This depends a lot on the performance level's you are expecting. In many
>> cases, native performance > VM performance.
>>
> That's something worth exploring actually.

I did say 'many cases' not 'all cases'. There are plenty of scenarios 
where VM's can actually run faster than native! One place where Virt 
falls quite hard on its face compared to native is on high load db 
traffic, specially when storage is local.

> A couple of years ago IBM produced a white paper where they measured
> infiniband connected machines as having performance within I thin 5%
> of native machines doing the same computation. I'm not sure what the
> current state of this is.

I doubt it would be far off that ..

> Also for desktops Intel's VT-d gives direct access to a netork card
> and a graphics card from a virtual machine, which is very interesting.

That is interesting, an AMD64 based test I ran late last year showed 
para-virt network performance to be much better than HVM on Xen. I've 
not done any such testing with intel kit though.


-- 
Karanbir Singh : http://www.karan.org/  : 2522219 at icq
-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug




More information about the GLLUG mailing list