RE: [Xen-users] Xen with Multiple NIC as a logical router

["Fischer, Anna" <anna.fischer@xxxxxx>]

> -----Original Message-----
> From: xen-research-bounces@xxxxxxxxxxxxxxxxxxx [mailto:xen-research-
> bounces@xxxxxxxxxxxxxxxxxxx] On Behalf Of Dimitrios Kalogeras
> Sent: 20 June 2008 12:39
> To: xen-users@xxxxxxxxxxxxxxxxxxx
> Cc: xen-research@xxxxxxxxxxxxxxxxxxx
> Subject: [Xen-research] Xen with Multiple NIC as a logical router
>
> Hi *,
>
> My sincere apologies for cross posting in the mailing lists.
>
> I am evaluating the case of using XEN as a logical router.
>
> http://www.cs.ucl.ac.uk/staff/M.Handley/papers/xen-vrouters.pdf
>
> In the conclusions section it is mentioned
>
> "  CPU saturation is a main feature of PC-based virtual routers, and
> that to avoid context switching overheads, the virtualization platform
> (e.g. Xen) so all forwarding is handled in the privileged domain. In
> other words, domUs should only host the control (slow) path of its
> associated virtual router, while the corresponding forwarding
> should be âmigratedâ to dom0. "
>
>
> I am considering the case to provide every DOMU a
> separate physical port. I know that this solution does not scale but do
> you believe that such a choice would limit the CPU overhead caused by
> the copying functionality between the DOMU and DOM0 ?

Sure, there are various ways to improve network performance and limit CPU 
overhead for DomUs. One would be to assign dedicated physical NICs to guest 
domains via pass-through as you suggest but the big problem is that you need a 
physical NIC per guest which is as you say not scaling very well. The next 
approach is to use virtualization-aware NICs that actually provide VNICs in 
hardware (PCI-SIG IOV). With that you can concurrently and directly access a 
single card from multiple guests. A third approach is to improve the 
performance of the inter-domain communication channels between DomU and Dom0 in 
Xen to reduce the CPU overhead for guest domain networking. There's currently 
work going on in all of those areas.

Regarding the virtual router approach there have been other reasons to leave 
the forwarding paths within a single domain instead of distributing them across 
multiple guests - those are also application-specific and not just performance 
related.

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