[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [Xen-users] SMP support in Windows domU


  • To: bdubas@xxxxxxxxxxx, xen-users@xxxxxxxxxxxxxxxxxxx
  • From: "Petersson, Mats" <Mats.Petersson@xxxxxxx>
  • Date: Mon, 14 May 2007 11:07:44 +0200
  • Delivery-date: Mon, 14 May 2007 02:06:39 -0700
  • List-id: Xen user discussion <xen-users.lists.xensource.com>
  • Thread-index: AceV/hj4WXhQR7MgSruXV9A+zOhZ3gAB+XSA
  • Thread-topic: [Xen-users] SMP support in Windows domU

 

> -----Original Message-----
> From: xen-users-bounces@xxxxxxxxxxxxxxxxxxx 
> [mailto:xen-users-bounces@xxxxxxxxxxxxxxxxxxx] On Behalf Of 
> bdubas@xxxxxxxxxxx
> Sent: 14 May 2007 09:01
> To: xen-users@xxxxxxxxxxxxxxxxxxx
> Subject: RE: [Xen-users] SMP support in Windows domU
> 
> Hi,
> 
> At last I can notice that I run Windows 2003 in SMP mode on top of XEN
> 3.0.4 that is currently available on Gentoo 2007 profile. I have some
> minors problems with performance witch is:
> 1. Network use some CPU while I transfer big files (probably 
> it is related
> to not using hardware CRC computing). Maximum transfer that I 
> reached is
> some about 10mbps with 30% of processor usage.

Yes, that's probably about par for the course. It will vary a little bit
based on which network device type you use - but not much. 

The right solution is to install a Para-virtual driver, but for Windows,
you can't just build one yourself (unless you are experienced in Windows
driver design), so you need one that matches your Xen version (easiest
way to achieve that is to install a package with Xen and the PV driver,
such as XenExpress). The para-virtual driver will install as a device
driver in Windows, but instead of talking to a virtualized "true"
network device, it just packs up the data to go on the network and sends
it directly to the Dom0 network backend driver. 

The reason that you get poor performance without the PV driver is that
the device-modeling is adding processor usage to the processing of the
data going to the network, because each access to the network device
will be intercepted by the hypervisor, sent off to Qemu-dm in Dom0, and
then processed there, before the hypervisor re-awakens the guest. This
is probably about 10x slower than doing a PCI-access on real hardware. 

> 2. When I install software lots of processor time is used by 
> the kernel
> mode (this is probably related to disk IO operations). It use 
> about 90% of
> processor while the application (user processor time) use about 10%.

Yes, that's probably correct. Again, you can get around this with PV
drivers as described above (but now transferring Disk-packets to Dom0). 

> 3. When I use more than 2 processors for HVM quest, after it 
> start XEN top
> show that it use 300% (for 4 processors) even if it don't perform any
> operation. Inside the quest manager show that system use about 6% of
> processor resources.

This doesn't sound right to me. I'm not sure what could be wrong tho'. 

--
Mats
> 
> Any way in the normal work when all is installed the 
> virtualization work
> very stabile and the performance its enough for what we do, 
> but it will be
> good if it could be improved somehow. So if somebody have 
> some experience
> and suggestion how to improve it I will be gratefully.
> 
> Best Bart.
> 
> 
> _______________________________________________
> Xen-users mailing list
> Xen-users@xxxxxxxxxxxxxxxxxxx
> http://lists.xensource.com/xen-users
> 
> 
> 



_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.