Lothar
*Von:*Joe Touch [mailto:[email protected]]
*Gesendet:* Dienstag, 16. August 2016 01:07
*An:* Reith, Lothar <[email protected]>; David Allan I
<[email protected]>
*Cc:* Bocci, Matthew (Nokia - GB) <[email protected]>; [email protected]
*Betreff:* Re: AW: AW: [nvo3] FW: Call for interest on NVO3 use case draft
On 8/15/2016 3:54 PM, Reith, Lothar wrote:
Hi Joe,
thanks for the references to these RFCs, in particular to RFC 3819
which I was not aware of.
I think the references make very clear, that the term “subnet” is
overloaded and therefore context dependent in meaning.
It's almost always more useful to be clear about the way in which the
term is used, but IMO "subnet" is an L3 term that refers to
addressing. Everything else derives from that.
I do not want to bother you with a detailed analysis, but my
observation is as follows:
The meaning of the term subnet is context dependent
- Layer-2 subnet in the sense of an object that forwards packets
- Layer-3 subnet in the sense of an object that forwards packets
- network prefix in the sense of an object that is a portion of an
address plan, more precisely a subtree of a binary tree.
L2 doesn't have subnets; a specific L2 *is* a subnet to L3.
L3 subnets are defined by having a common bit-aligned address prefix.
Adding to this ambiguity is the fact that subnet number (IPv4)
and subnet ID (IPv6) relate to the Layer-3 subnet context of
“subnet” only
That's because subnet is an L3 term.
(not to the layer 2 subnet context)
There is no L2 subnet. An L3 subnet is often - but not always - mapped
to a single L2 *network*.
but do not identify a layer-3 subnet as a global identifier,
rather play the role of a locator in the binary subtree with local
significance only. This appears to be similar to the locator ID
separation problem addressed by LISP.
That's exactly because LISP turns the encapsulation layer into what is
effectively an L2 network. The problem of mapping the inner and outer
addresses in LISP is the same as ARP, except that it requires
determining LISP egress rather than destination.
That's why we called the protocol to solve this "BARP" - it combines
BGP and ARP - when we developed it for what we called "recursive
routers" in the late 90s in our X-Bone system.
Given this and Dave’s additional comments regarding: “please
provide a use case please” may I kindly ask you to provide a use
case with “context tag” or perhaps using the term network prefix
instead of subnet where appropriate.
In particular I would be interested in a use case describing your
stated requirement for “revisitation, where a single node
participates multiple times in an overlay”. Could you please
clarify the cardinality relations in this use case, e.g. do you mean
- One physical interface to appear like multiple physical
interfaces (such as a single physical NIC to appear as multiple
VNICs)?
- One physical interface (typically Ethernet Interface or
station) to be associated with one MAC addresses at Layer 2,
where multiple IP addresses are associated with that MAC address,
where said IP addresses belong to multiple Layer3 subnets., thus
multiple Layer3 subnets associated with the same Layer2 subnet?
- One physical interface (typically Ethernet Interface or
station) to be associated with multiple MAC addresses at Layer 2,
where each of that MAC addresses is assigned one IP address (one
to one relation between Layer3 subnet and Layer2 subnet, but
multiple layer2 subnets on the same physical interface (Ethernet
station i.e. PNIC or VNIC)?
First, if you're thinking about physical anything, you're limiting
yourself to one layer of virtualization. That's unnecessary.
So let's assume that when you say "physical" you really mean "in the
lowest layer of virtualization" (i.e., the base case of what is
ultimately a recursive structure - e.g., see www.isi.edu/rna
<http://www.isi.edu/rna>, which is a generalization of the X-Bone
architecture).
When I say "revisitation", I mean one base L3 interface that acts like
multiple NVO3 L3 interfaces.
That is accomplished by associating it with multiple virtual L2 (L2
over L3), each virtual-L2 of which is associated with one virtual L3.
None if this is what you're describing above; you're stuck at the base
layer L2, which is below where NVO3 operates.
Joe
Best Regards, Lothar
PS: Given that we live in times of emerging software defined
networks, it is increasingly important WHAT an orchestration
software defines when it defines a “network” or a “subnet”, and
rapidly becoming irrelevant what context tag was not explicitly
spelled out in an IETF RFC. So I recommend to analyse WHAT it is,
that the OpenStack method “create network” creates. It may be an
empty binding between a L2-subnet and a L3-subnet with all details
and sizes yet undefined, but with a handle in form of a globally
unique identifier in the form of a UUID, and the problem solved is
the lack of a global identifier that is not a locator or address.
*Von:*Joe Touch [mailto:[email protected]]
*Gesendet:* Sonntag, 14. August 2016 01:45
*An:* Reith, Lothar <[email protected]>
<mailto:[email protected]>; David Allan I
<[email protected]> <mailto:[email protected]>
*Cc:* Bocci, Matthew (Nokia - GB) <[email protected]>
<mailto:[email protected]>; [email protected] <mailto:[email protected]>
*Betreff:* Re: AW: [nvo3] FW: Call for interest on NVO3 use case draft
On 8/13/2016 4:20 PM, Reith, Lothar wrote:
Hi Joe, Dave, dear all,
I think that the term “subnet” is indeed only defined in a context
sensitive way.
Unfortunately there is no distinction made between a L2-subnet
(aka a broadcast domain, or a MEF EVC, or o broadcast domain
equivalent achieved without broadcast such as EVPN) and a
L3-subnet (the object that does not appear in any information
model that I am aware of, but which has a size which is defined
by a subnet-mask and which supports the use case of “change the
IP-address range assigned to that subnet”.
The reason that L2 broadcast domain and L3 subnet (as per RFC1812)
coincide is that is the basis of most L3:L2 mapping mechanisms,
e.g., ARP (RFC826) or ARP emulation (RFC1577).
And the concept of an L3 subnet is pervasive in many RFCs, being
the basis of hierarchical IP forwarding since its inception.
There's even a specific RFC advising on the design of L2 subnets
to support L3 subnets (3819).
Part of the reason for complexity is the lack of a proper
definition of this object in the information model.
If I am wrong – I would be happy to be corrected by someone who
can point me to an authoritative definition of the term subnet as
an object that supports the use case “change address range
assigned to subnet” without specifying how the subnet is built as
layer 2 construct (e.g. as Ethernet yellow cable, VLAN in a
Switched Ethernet domain, VXLAN overlay, EVPN, MEF EVC or ATM
point to point link).
Subnet in L3 in IPv4 is defined as per RFC 4632
Subnet in L3 in IPv6 is defined as per RFC 4291
These are very old and fundamental concepts to the Internet
architecture. Any notion of virtualizing L3 needs to deal with them.
Joe
Thanks to OpenStack for finally fixing the problem by introducing
the method “create network”, which creates – exactly – this
missing object in an abstract way.
So Dave’s question is very valid, simply because the term
“subnetting” is not properly defined – unless the authors point to
a reference RFC where the term is defined in an authoritative way.
Lothar
*Von:*nvo3 [mailto:[email protected]] *Im Auftrag von *Joe Touch
*Gesendet:* Samstag, 13. August 2016 19:30
*An:* David Allan I <[email protected]>
<mailto:[email protected]>
*Cc:* Bocci, Matthew (Nokia - GB) <[email protected]>
<mailto:[email protected]>; [email protected] <mailto:[email protected]>
*Betreff:* Re: [nvo3] FW: Call for interest on NVO3 use case draft
On 8/13/2016 9:52 AM, David Allan I wrote:
Hi Joe
And the use case for wanting to do subnet emulation is….?
You want the properties of a subnet and/or to emulate the behavior
of a shared link, i.e., to limit the scope of various protocols,
including IP routing, IPv6 automatic addressing, L2 address
translation (virtualizing L2 underneath a virtual L3 is needed to
support revisitation, where a single node participates multiple
times in an overlay), and basically any subnet-based resource
discovery.
Joe
That‘s my question
Dave
*From:*Joe Touch [mailto:[email protected]]
*Sent:* Friday, August 12, 2016 8:20 PM
*To:* David Allan I <[email protected]>
<mailto:[email protected]>
*Cc:* [email protected] <mailto:[email protected]>; Bocci, Matthew (Nokia
- GB) <[email protected]> <mailto:[email protected]>
*Subject:* Re: [nvo3] FW: Call for interest on NVO3 use case draft
The typical use case is to support subnet emulation, e.g., a group
of links over which broadcast is emulated as with LANE.
On Aug 12, 2016, at 7:11 PM, David Allan I
<[email protected] <mailto:[email protected]>>
wrote:
My point would be that introducing additional complexity in an
overlay should have a use case associate with it. It would not be
something you would do gratuitously….
SO I’m looking for the draft to provide a use case for this vs.
simply mentioning subnetting without any context J
Cheers
Dave
*From:*nvo3 [mailto:[email protected]] *On Behalf Of *Joe Touch
*Sent:* Friday, August 12, 2016 5:07 PM
*To:* David Allan I <[email protected]
<mailto:[email protected]>>; [email protected]
<mailto:[email protected]>; Bocci, Matthew (Nokia - GB)
<[email protected] <mailto:[email protected]>>
*Subject:* Re: [nvo3] FW: Call for interest on NVO3 use case draft
On 8/12/2016 4:16 PM, David Allan I wrote:
4.2 Why I would subnet my overlay could use some explanation. I
normally think of subnetting as a convenient address
summarization technique dependent on topology, and with an overlay
I don’t have a topology.
The topology of an overlay is determined by its tunnels, just as
the topology of the underlying net is determined by its links.
A subnet in an overlay corresponds either to a single multipoint
tunnel or to a set of tunnels that transparently acts as such -
just as a subnet in the Internet base network corresponds to a
shared access link or a set of links that transparently act as
such (e.g., switched ethernet).
Joe
