Re: [vnrg] way forward on VNRG definitions

[Joe Touch <[e-mail]>]

Hi, all,

Speaking as an *individual* contributor (i.e., with my RG-chair hat "off"):

On 6/8/2010 9:57 AM, Joe Touch wrote:
> -----------------------------------------------------
>
> Starting questions:
>
> 1. how do you define VNs?

I define them as composed of virtual routers, virtual hosts, and virtual 
links.

A virtual link is the easiest to define - it is a tunnel over some 
existing network path, i.e., with an additional layer of encapsulation 
that is used solely for the VN, but which is otherwise not needed.

A virtual host is a source or sink of packets on a virtual link.

A virtual router forwards packets between virtual links.

Note that nothing about these definitions specifies a boundary, i.e., 
inside a single machine, etc. I don't think those boundaries are 
meaningful in the base Internet anyway.

I.e., IMO, virtual has nothing per se to do with "logical". I.e., a set 
of devices on a network that source/sink packets with a single network 
address act as a single logical host. That's not 'virtual' to me, though 
most 'virtual' things tend to be logical, not all logical things are 
virtual.

> 	1.a. what are the key components?

See above; virtual link, virtual router, and virtual host.

> 	1.b. what is the relationship between these components?

See above; hosts source/sink packets, and routers forward them. When a 
router sources/sinks packets, IMO it acts as a host on the network for 
those messages.

> 	1.c. what is the characteristic behavior/capability of the
> 	resulting system?

I think that the defining characteristic of a VN is that it allows an 
existing network to concurrently emulate another, distinct and separate 
network.

> 2. what are VNs used for?

Protection: keeping concurrent uses of a network from interfering with 
each other.

Abstraction: presenting the application (whether user layer, routing, 
etc.) with a simpler or desired topology that need not match the 
physical connections. E.g., as with DHTs, hypercubes, rings, etc. This 
also includes presenting a topology that is either much larger or 
smaller than the physical, i.e., that doesn't match in scale either.

Sharing: allowing a single set of resources to be concurrently used for 
multiple, separate purposes.

NB: these are identical to the reasons why virtual memory is useful

> 3. what are they key challenges?
 > for each challenge:
 > 	- define the challenge
 > 	- explain why it is hard
 > 	- provide some references to those working on solutions

Challenge: mapping logical components onto virtual ones - which I 
consider the 'provisioning' step of a conventional physical network.

Why it's hard: like any mapping, this is a distributed resource 
allocation problem, difficult to optimize fully, and challenging to 
coordinate without requiring global locks.

Who's working on this: this is known as "mapping" VNs, and many groups 
have worked on it over the years; it's also one of the least interesting 
problems, IMO, since it is not particularly new in the virtual domain 
(i.e., it exists in physical networks too)

--

Challenge: Clean, clear architecture supporting VNs.

Why it's hard: done wrong, VNs require users to reimplement some 
protocols at the user layer (e.g., as with DHTs, DTNs, etc. - usually 
transport protocols). done wrong, VNs also require extraordinary 
measures to restore existing capabilities, such as traceroute, ping, etc.

Who's working on this: our group (X-Bone), Violin, Akari (IMO)

--