Re: [vwrap] VWRAP and IPv6

[Morgaine <[e-mail]>]

Abstracting over addresses is fine, Richard, but after you've dealt with the
abstractions you typically end up with addresses, and that's what protocols
work with from then on.

The textual abstractions are only there for human comfort, ease of
configuration, and great flexibility of indirection, but abstractions are
just means to an end.  After the address abstractions are de-abstracted,
protocols work with the much more efficient binary representations.  That's
practically universal.

It's easy to give a practical example of such operation.  For example,
consider a design in which all VW objects have two parts, an in-world object
(IWO) and an in-client object (ICO), where the ICO is replicated in each
client.  Let both parts be active, ie. both sides can be event generators.
It's easy for ICOs to address update events to their IWO because there is
only one IWO, and access could be optimized to network level after the
de-abstraction step for efficiency.  But what about updates in the reverse
direction, which is 1:N?

In the typical IPv4 environment, the usual approach would be a heavily
indirected one, with updates travelling through layer upon layer of
middleware before reaching their target ICO.  In an IPv6 environment in
which each ICO can have its own IPv6 address, there need be no middleware at
all between the host kernel and the target object.  The resulting increase
in efficiency by orders of magnitude can have major ramifications on the
operation of your world and of your clients.

And the cost of this massive leap in efficiency?  Merely to inform the IWO
of the ICO's dynamic IPv6 network address, which borders on the trivial.

One of my main interests is VW scalability, both of worlds and of clients,
and such efficiency gains are truly important.  Scaling VW complexity and
rendering of crowds or of highly detailed visualizations requires some
lateral thinking and more powerful networking paradigms than we have been
used to until now.  IPv6 offers some interesting solutions in that
direction, and they will be used.


Morgaine.





==================================

On Thu, May 20, 2010 at 2:43 PM, Richard Barnes <[e-mail]> wrote:

> As an aside, there's nothing that says that subscribers have to be
> allocated a /64.  Some proposed addressing schemes that are more
> conservative (in the Asia/Pacific region IIRC) actually assign much longer
> prefixes (smaller address blocks) to customers.
>
> Fine-grained addressability is not the same as fine-grained allocation of
> *IP* addresses. Think of it this way: HTTP has very fine-grained
> addressability, with individual identifier/locators for pages, images,
> scripts, etc.  It's able to accomplish this not by assigning an IPv6 address
> to every object, but by using an application-layer identifier/locator --
> namely a URL.
>
> Not using IPv6 addresses in applications is not painting oneself into an
> IPv4 corner, it's good protocol design following the layering model of the
> Internet.  (If we're using IPv4 addresses for identifiers, we shouldn't be
> doing that either!)  Abstracting over IP addresses is about freeing the
> application from having to worry about whether its IPv4 or IPv6 or IPv10 or
> X.25 or whatever else at the network layer.
>
> --Richard
>
>
>
>
> On May 20, 2010, at 9:23 AM, Morgaine wrote:
>
>  On Thu, May 20, 2010 at 10:59 AM, Han Sontse <[e-mail]>
>> wrote:
>> I recognize that it is temping to take advantage of the huge address space
>> that IPv6 offers but my test engineer instincts tell me it is unwise to do
>> so.
>>
>>
>> I too am an engineer, but solid engineering instincts do not tell us to
>> use only those facilities whose numbers or sizes can be tested exhaustively.
>>  If it were so, we would not have multi-terabyte disk drives today.
>>
>> The /64 minimum routable subnet of IPv6 address space was not designed as
>> it was because people were expected to need 2^64 NIC cards in their PCs.
>>  This immense per-user address block was created because IPv6 designers
>> could see beyond IPv4's use of addresses for host interface identification,
>> and were visionary enough to realize that all objects could benefit from
>> network addressability.  This also happens to be one of the key messages of
>> REST.  Addressability is extremely important, the finer the better, not only
>> on the Web, and not only with URIs.
>>
>> With IPv6 just around the corner, I foresee such addressability at fine
>> levels of granularity appearing in VWs relatively soon.  IPv4 worlds will
>> have to emulate what IPv6 worlds achieve natively, and that is the right way
>> around.  When legacy design restricts progress, the only thing it achieves
>> is rapid obsolescence.
>>
>> I would prefer VWRAP not to be painting itself into an IPv4 corner of
>> obsolescence before it's even released, which is why I'm raising this topic
>> --- it's intended as a heads-up for the group.
>>
>> Since the rise of VWs and IPv6 take-up are going to occur concurrently, I
>> think the next few years are going to be very interesting on the VW
>> networking front.  My recommendation is to observe carefully what happens,
>> and not to make any rash decisions based on legacy ideas that could cut off
>> your future.
>>
>>
>> Morgaine.
>>
>>
>>
>>
>>
>> ===================================
>>
>> On Thu, May 20, 2010 at 10:59 AM, Han Sontse <[e-mail]>
>> wrote:
>> It's a bad idea to use IP addresses for anything other than addressing a
>> machine level interface.  Using them to interface objects at or above the
>> application level is asking for trouble.   UUIDs are UUIDs... IP addresses
>> are IP addresses.   The two should not be mixed.   I recognize that it is
>> temping to take advantage of the huge address space that IPv6 offers but my
>> test engineer instincts tell me it is unwise to do so.
>>
>> Han.
>>
>> On May 20, 2010, at 2:50 AM, Morgaine wrote:
>>
>>  A /64 is the smallest routable IPv6 subnet, hence there is no extra
>>> routing overhead for the 2^(128-64) addresses in the basic block that every
>>> user will have allocated.
>>>
>>> Needless to say, there won't be anything like 2^64 objects in your
>>> scenegraph, so if you use your IPv6 address space to address every
>>> scenegraph object that you have instantiated client-side, you will still
>>> have an extremely sparsely used /64 block.  You'll be keeping track of which
>>> addresses you have assigned dynamically of course. (An address space of this
>>> size is not linearly scannable in realistic time.)
>>>
>>> In principle this is no different to using UUIDs --- your IPv6
>>> micro-servers will have unguessable IPv6 addresses within your /64 block,
>>> and it'll be up to you to tell other participants in the protocol (for
>>> example server-side objects) with which of your IPv6 micro-servers they are
>>> to communicate.  This is why an ADT that is supportive of IPv6 operation
>>> needs to be able to carry such data efficiently --- the UUID is no longer
>>> the only 128-bit datum of interest.
>>>
>>> Of course your entire /64 block is routed as one, since there are no
>>> smaller routable subnets in IPv6, and therefore your objections about
>>> routing non-scalability do not apply.
>>>
>>>
>>> Morgaine.
>>>
>>>
>>>
>>>
>>> =====================================
>>>
>>> On Wed, May 19, 2010 at 10:56 PM, Richard Barnes <[e-mail]>
>>> wrote:
>>> The impact of IPv6 on VWs should be far less than you imagine, I think.
>>> :)
>>>
>>> It has often been said that many, many things will be assigned IPv6
>>> addresses.  There will certainly be enough to go around, but that doesn't
>>> make them good application-layer identifiers.  If you want to use IP
>>> addresses with IP, they have to route, and that means that they have to be
>>> either (1) topologically assigned, or (2) overlay routed, e.g., with mobile
>>> IP.  So if you want to use an IP address to address something, you're going
>>> to have to either change that address when the thing gets connected
>>> differently, or you'll need some overlay routing scheme to route to a
>>> persistent address.  (And if you're not going to use the address with IP,
>>> then why would you use an IP address?)
>>>
>>> VWRAP, as an application layer protocol, has a choice about what
>>> identifiers to use for things and how to use those identifiers.  ISTM that
>>> using IP addresses (again, of either type) remains a bad idea, compared with
>>> something that is decoupled from the routing layer.
>>>
>>> As far as the DNS, I'm not sure what your point is:
>>> (i) DNS caching has been around for decades
>>> (ii) DNS names can already refer to multiple addresses
>>> (iii) Depends on implementation.  Maybe there's some application-layer
>>> way to resolve names.  E.g., SIP endpoints don't need to have DNS records
>>> because the INVITE transaction maps from SIP URIs to IP addresses.
>>> (iv) From RFC 4291: "IPv6 addresses are 128-bit identifiers for
>>> interfaces and sets of
>>> interfaces"
>>> (v) Seems like there's a use case lurking, hard to address without
>>> details.  But you're probably still at the application layer.  For instance,
>>> if you were telling your peer to send you an HTTP request, you would send an
>>> HTTP URI, which can already handle IPv6 addresses.
>>>
>>> --Richard
>>>
>>>
>>>
>>> On May 19, 2010, at 9:35 AM, Morgaine wrote:
>>>
>>> The impact of IPv6 on VWs is going to be far greater than you imagine, I
>>> believe.  As things stand, VWRAP is being designed around the features and
>>> limitations of IPv4 networking.  It will of course be able to ride on top of
>>> IPv6 as a transport, but unable to benefit from its improved feature set,
>>> unless we provide the means to do so.
>>>
>>> The question "Why is there ever a need to carry an IP address in this
>>> application-layer protocol, as opposed to, say, a DNS name or a URI?"
>>> suggests an IPv4 mindset, because in IPv6 it has a variety of possible
>>> answers:  (i) When the DNS lookup has already been done and does not need to
>>> be done again, (ii) When a FQDN maps to multiple addresses, but you want to
>>> refer to a specific one, (iii) When an IPv6 address does not have an
>>> external address record, (iv) When the host is already known and the IP
>>> address denotes something other than a physical NIC interface, or (v) When
>>> listeners are allocated locally and dynamically for the current session
>>> only, and their IP addresses need to be communicated to a peer --- this will
>>> be an extremely common use case with IPv6.
>>>
>>> In today's IPv4-based systems, it is normal for a user's machine to have
>>> a single IP address, and for a client on that machine to talk to a server
>>> that has a single customer-facing IP address too (although it may also have
>>> others that are not public).  In IPv6-based systems, this simple scenario is
>>> still possible but represents only a special case among a new and vastly
>>> expanded set of IPv6 use cases.
>>>
>>> The smallest allocatable block of IPv6 address space is a /64, so
>>> 2^(128-64) IP addresses is the number that most normal IPv6 users will have,
>>> although the much larger /48 blocks of 2^80 addresses are commonly assigned
>>> to users as well.  With such IP riches client-side, it is certain that they
>>> will be used for something, and the obvious suggestion is to give every VW
>>> object an IPv6 address of its own.
>>>
>>> Lest we forget, one of the most important properties of the REST paradigm
>>> is addressability, but the benefits of high addressability go far beyond the
>>> Web.  This is a property that will become ubiquitous throughout networked
>>> computing with IPv6.  It's probably no exaggeration to say that *everything*
>>> will be getting its own IP address. :-)
>>>
>>> Faced with that future, it's a mistake to think in IPv4 terms and
>>> consider IPv6 addresses as merely denoting a host which can be obtained from
>>> its FQDN.  That's IPv4 thinking, and represents only a very basic use case
>>> for IPv6.  In more extended use cases, IPv6 addresses will become much more
>>> like active local UUIDs than like today's host IP addresses.  They will be
>>> used for inner-application addressing, for example to make every object in
>>> an application an addressable micro-server.
>>>
>>> VWRAP is being born into an IPv6 world, and it would be a mistake to
>>> hardwire IPv4 thinking into its design.  We continually refer to the need
>>> for extensibility and a desire to future-proof the protocol, but lip service
>>> is not enough.  Adding IPv6 addresses to the underlying ADT would provide a
>>> bare minimum level of thinking ahead to support VWRAP implementations that
>>> embrace large numbers of communicating objects.  (A generally extensible ADT
>>> would of course achieve the same thing, and more.)
>>>
>>>
>>> Morgaine.
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> ==============================================
>>>
>>> On Mon, May 17, 2010 at 3:50 PM, David W Levine <[e-mail]> wrote:
>>>
>>>
>>> [e-mail] wrote on 05/17/2010 10:08:26 AM:
>>>
>>> > [image removed]
>>> >
>>> > Re: [vwrap] VWRAP and IPv6
>>> >
>>> > Richard Barnes
>>> >
>>> > to:
>>> >
>>> > Morgaine
>>> >
>>> > 05/17/2010 10:09 AM
>>> >
>>> > Sent by:
>>> >
>>> > [e-mail]
>>> >
>>> > Cc:
>>> >
>>> > vwrap
>>> >
>>> > In the spirit of good layering, it seems like we should be talking
>>> > about avoiding direct reliance on IP addresses instead of how we can
>>> > accommodate different addresses in the protocol.  Why is there ever a
>>> > need to carry an IP address in this application-layer protocol, as
>>> > opposed to, say, a DNS name or a URI?
>>> >
>>> > --Richard
>>> >
>>> +1
>>>
>>> We may from time to time have a IP address stuck in the form of a URI
>>> which has the dotted decimal form of an resource. The spec should make
>>> sure that form can be an IPV6 address, but that should be very straight
>>> forward.
>>>
>>> There may be some IPV6 features which eventually impact us at the
>>> protocol
>>> design level, but the entire VWRAP specification has been written in
>>> terms of
>>> URIs and Capabilities, not in terms of low level IP addresses. It would
>>> take an
>>> incredibly powerful argument to suggest that this should change.
>>> Likewise, while
>>> IPV6 changes addressing at the bottom of the stack, it doesn't
>>> appreciably change
>>> the cost of holding open sessions, of managing sockets and buffers inside
>>> of
>>> clients (especially lightweight devices) Further while we can all
>>> devoutly hope
>>> that IPV6 finally deploys broadly, making it a dependancy in the near
>>> term feels
>>> rather like a roadblock to possible adoption.
>>>
>>> - David
>>> ~ Zha
>>>
>>>
>>>
>>> >
>>> > On May 14, 2010, at 7:32 AM, Morgaine wrote:
>>> >
>>> > > Despite exceedingly low takeup of IPv6 so far, it's coming, and in
>>> > > the context of IETF WG rates of progress, it is really imminent.  Of
>>> > > special relevance to us is that IPv4 addresses are projected to run
>>> > > out within roughly the same time frame as VWRAP materializes.
>>> > >
>>> > > There is little point in creating a new VW protocol today purely in
>>> > > the context of IPv4 without thinking about tomorrow's IPv6
>>> > > environment.  By the time that VWRAP is fully defined, we should
>>> > > expect a substantial and rapidly increasing proportion of VWRAP
>>> > > implementations to occur in IPv6-accessible worlds.  This is a
>>> > > situation for which we need to plan now, otherwise the protocol will
>>> > > be obsolete before it's even ready.
>>> > >
>>> > > To get the ball rolling, here are a few IPv6-related issues that we
>>> > > could usefully address in the group:
>>> > >
>>> > >    • The 128-bit IP address of IPv6 is likely to become a common
>>> > > payload, so it should be supported natively.  LLSD already supports
>>> > > one 128-bit data type of course, the UUID.  IPv6 addresses should
>>> > > not be transported in UUID fields however, since their statistical
>>> > > properties and semantics are totally different.  In any case, the
>>> > > two spaces are disjoint so confusing them would be a mistake.  The
>>> > > IPv6 address seems a good candidate for its own defined type, in the
>>> > > absence of wide integers.
>>> > >    • The extremely large address space of IPv6 has often been
>>> > > portrayed as allowing everything around us to have its own IP
>>> > > address.  In the context of virtual worlds, this idea can actually
>>> > > be put to good use, with each virtual object bearing its own IPv6
>>> > > address.  One candidate application for this has already been
>>> > > described in earlier discussions --- for example, implementing
>>> > > virtual object simulation in two parts, one client-side and one
>>> > > world-side, the two parts communicating directly.  IPv6 addressing
>>> > > of objects makes this relatively easy to achieve efficiently, so its
>>> > > implementation in IPv6-based worlds is almost inevitable before
>>> > > long.  VWRAP should be ready for this.
>>> > >    • The event queue in current implementations provides an indirect
>>> > > and highly inefficient way of achieving world-to-client
>>> > > communications.  IPv6 offers opportunities for direct communication
>>> > > that can make the event queue approach obsolete, so it's worth
>>> > > examining IPv6-oriented alternatives.
>>> > >    • We recently discussed feature and protocol negotiation within
>>> > > VWRAP.  IPv6 provides much opportunity for new features and protocol
>>> > > pathways to be negotiated between endpoints, so it gives us a very
>>> > > real and imminent set of use cases for negotiation.
>>> > >
>>> > > It would also be a feather in VWRAP's cap to be considered an "IPv6-
>>> > > capable application protocol" in a stronger sense than just to be
>>> > > carried by IPv6 networks.  IPv6 does after all bring new abilities
>>> > > to the table, such as the huge addressing, and it would be a mistake
>>> > > to ignore them.
>>> > >
>>> > >
>>> > > Morgaine.
>>> > >
>>> > >
>>> > >
>>> > > _______________________________________________
>>> > > vwrap mailing list
>>> > > [e-mail]
>>> > > https://www.ietf.org/mailman/listinfo/vwrap
>>> >
>>> > _______________________________________________
>>> > vwrap mailing list
>>> > [e-mail]
>>> > https://www.ietf.org/mailman/listinfo/vwrap
>>>
>>> _______________________________________________
>>> vwrap mailing list
>>> [e-mail]
>>> https://www.ietf.org/mailman/listinfo/vwrap
>>>
>>> _______________________________________________
>>> vwrap mailing list
>>> [e-mail]
>>> https://www.ietf.org/mailman/listinfo/vwrap
>>>
>>> _______________________________________________
>>> vwrap mailing list
>>> [e-mail]
>>> https://www.ietf.org/mailman/listinfo/vwrap
>>>
>>
>>
>> _______________________________________________
>> vwrap mailing list
>> [e-mail]
>> https://www.ietf.org/mailman/listinfo/vwrap
>>
>
>