Re: [nvrg-bof] What is Network Virtualization?

[Joe Touch <touch at ISI.EDU>]

Andreas Wundsam wrote:
> Hi,
>
>
> Joe Touch schrieb:
> > Nick Feamster wrote:
> > > I agree. unneling is a *mechanism* by which links can be instantiated
> > > over a shared physical infrastructure, but the question "What is
> > > network virtualization?" is really a what question (tunneling answers
> > > "How?", not "What?") 
> > If that's true, then why is the Internet not already a virtual net, 
> > e.g., supporting different transport protocols over IP? IP is shared 
> > by different transports. Similarly, ethernet supports different 
> > network protocols.
> I would think think that the key difference is the level of transparency 
> and information hiding offered. Virtualization generally is transparent 
> to the user, layering is not. TCP knows it is running on top of IP. A 
> network stack running in a virtualized network would believe itself to 
> be running alone on a dedicated physical infrastructure, when in fact it 
> could coexist with other networks on the same hardware. The analogy with 
> Virtual Memory holds here - a program running on virtual memory does not 
> have to know other processes are existing on the system -- all it sees 
> is a flat address space starting at 0. Hence, I would consider VLANs, 
> Tunnels etc. all proper mechanisms for virtualizing links, but only in 
> combination with a mechanism that hides their existence and specifics 
> from the user.
VLANs don't qualify by your requirements above; VLANs don't own a 
particular fraction of a link, so they could detect other VLANs on the 
same link. Further, VLANs virtualize the link, but not the endpoints - 
e.g., to have two segments with the same destination address and 
different VLANs still go to the same physical destination.
As to layering, TCP knows it's running on IP, but not that IP is running 
on ethernet (or, in fact, what it's running on). Ignorance of 
non-adjacent layers is a already a property of layering.
Support for separate stacks would include virtual hosts alone as 
sufficient to make a virtual network (e.g., running two OS's on the same 
hardware, which I don't think makes sense.
I still think VNs need a layer of indirection; there are plenty of 
interesting things that don't, and things that support VNs that don't, 
but they're not key to the concept. It's that layer of indirection that 
I describe as a tunnel - there are many ways to accomplish that, 
including shim headers (HIP, SHIM6) as well as encapsulation.
Joe

Attachment: signature.asc
Description: OpenPGP digital signature