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

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



Dear Joe,
- why is the router to be a virtual one? I understand a virtual router as a virtual machine which certainly can manage a real network. Since the router does not change anything, that the additional indirection information must initially be understood in some real manner, why would the routers have to different?

I suppose this may result from a vision of a network map like in a NIMROD with the information attached to the nodes. But if you consider that the information are attached to the virtual adjencies you do not need any node related information. You can compute presence, throughput and delays at related adjencies ends. This is more accurate as they may depend on the machine dynamic allocation to the VN resources for that node. The interest is that the difference between the I/O of one way of a virtual adjency gives the line entropy which may be more important to VN path optimization than node load?

- Interested to understand what difference you make between the added indirection information and presentation layer? My first guess is that they are the same except that you push a little more on what presentation is used to be considered in IETF ?

- any objection to have this indirection information made conditional in order to build dynamic VN? Depending on the router I come through I can select one routing or another one, or multiplicate (for example for security reasons or to support multicast or multihoming) ?

- can this indirection information also be virtual in your mind ? So I may use an addressing which is dedicated to the VN ? Or that can be dynamic.

- would you have an idea for a place where to locate that indirection information in packet headers ?

- one of the point I rose about multi-level addressing can also be discussed here.

You call a VH a place where the indirection information can be changed. What about "changed or should have been adequately changed". Let suppose the indirection information encapsulate a second level indirection. Now you have an OPES attached to the adjency (a shim at one of the I/O). If the indirection value the packet carries is the VH virtual address, the OPES can remove it, so the VH can directly act as a gateway using the new indirection?.

jfc

At 14:55 22/06/2008, Joe Touch wrote:
Based on exchanges with a few people on the list, I'll revisit my
initial proposal.

Some acid tests for a definition:
    - it should support VPN, PPVPN, and overlays as VNs
    - it should not define the native Internet as a VN
        i.e., it must distinguish layering from
        "overlayering"

Additional notes:
       - network virtualization describes the process of
       creating virtual networks, i.e.,
                NV is the process
                VNs are the artifact

I define NV and VNs via the artifact created, because NV is then most
generally defined:
        NV is *any* process or mechanism that creates (enables the
        creation) of VNs

A _VN_ is a network composed of virtual links, virtual hosts, and
virtual routers. Virtualization of links, hosts, and routers is
accomplished by adding a layer of indirection in the names and/or
addresses associated with each. A VL encodes this indirection sufficient
for use by VHs and VRs; VHs and VRs uses this indirection information to
associate with a VL. Additionally, a VH is a network node that adds or
removes indirection information, and is associated with at least one VL
in a given VN. A VR is a network node that does not add or remove
indirection information, and is associated with at least two VLs in a
given VN.

The rest, I beleive, remains a reasonable summary of the capabilities of
VNs:

---
Virtual networks have three primary uses:
    - protection
        allow new services/protocols to be deployed on a subset
            e.g., testbeds, incremental deployment
        keep experiments from leaking out
            e.g., testbeds
        keep others' uses from affecting a given use
            e.g., emergency services, guaranteed capacity,
            privacy/authentication
    - concurrency
        shared use of common infrastructure
    - abstraction
        simplify the topology (e.g., LISP/NERD)
        support application-specific topology (e.g., P2P)

I have not listed mechanisms that support VNs, i.e., NV mechanisms yet.
These might be summarized as:

NV mechanisms:

        - partitioning (as in Clonable Stacks in hosts/routers, and VPN
        IDs, VLAN IDs, and tunneling in links, etc.)

        - aggregation (channel bonding in links, NAT-like services
        at the edge of server farms and cluster computers, etc.)

        - combinations of the above

Joe








Note Well: Messages sent to this mailing list are the opinions of the senders and do not imply endorsement by the IETF.