Re: [MBONED] "MLDv2 Procedures for Link-Layer Unicast Delivery of Multicast"

[Mark Smith <[e-mail]>]

Hi Toerless,

Sorry not to get back to you sooner, a bit time constrained at the moment.

Thanks very much for your review and comments.

----- Original Message -----
> From: Toerless Eckert <[e-mail]>
> To: Mark Smith <[e-mail]>
> Cc: "[e-mail]" <[e-mail]>; Dave Thaler <[e-mail]>; "[e-mail]" <[e-mail]>; "[e-mail]" <[e-mail]>
> Sent: Tuesday, 2 April 2013 7:56 AM
> Subject: Re: [MBONED] "MLDv2 Procedures for Link-Layer Unicast Delivery of Multicast"
> 
> Bunch of comments...
> 
> 1. What Mark is proposing is being done AFAIK in a variety of products on the 
> market,
>    for example as the "multicast to unicast" eleement of Cisco 
> VideoStream 
>    and i think other vendors do this as well. Not sure about IPv6 though, 
> i've only seen
>    it in IPv4.
> 


That's good to know, I'd say it shows the idea is both feasible, useful and deployable.

I found the following, which was a reasonable overview:

http://www.cisco.com/en/US/prod/collateral/wireless/ps6302/ps8322/ps10315/ps10325/white_paper_c11-577721.html

Dale Carder also pointed off list that Aruba Networks do the same thing (under Dynamic Multicast Optimization):

http://www.arubanetworks.com/vrd/80211nnetworksvrd/wwhelp/wwhimpl/common/html/wwhelp.htm#href=Chap4_AdaptRadioMgmt.html&single=true

> 2. I am always getting annoyed when IPv6 bigots oops: fans unnecessarily 
> constraining
>    solutions to IPv6 only even though they equally apply to IPv4 and, in the 
> case
>    of IP multicast when its clear that the mayority of use for a while will 
> still be
>    IPv4. Its already bad enough that RFC6085 was written only for IPv6 instead 
> of
>    for IPv4 and IPv6. It would be lovely if Marks draft could be modified to 
> cover
>    both IPv4 and IPv6. I can't remember that RFC6085 was given for review to 
> MBoned,
>    or else i'd hae made that comment for it as well. But see 1: in practice 
> existing
>    IPv4 host stacks will wwell receive DMAC-unicast IP multicast packets due to 
> the
>    fundamental separation of L3 and L2 in the architecture of most host stacks.
> 

I initially limited the topic to IPv6, because I knew IPv6 had the functions necessary i.e. MLDv2 and IPv6 ND NUD. That being said, Carston has since pointed out that RFC4861 says that NUD is only supposed to be done for unicast traffic, but not multicast traffic. That would be talking about layer 3 unicast vs layer 3 multicast. To accommodate the use of NUD for what is proposed here, RFC4861 would have to be slightly updated to apply NUD to link-layer unicast of layer 3 multicast.

As for IPv4, that was something I was going to investigate. IGMPv3 would support this, the missing bit is an ARP equivalent of NUD. As far as I'm aware (although I haven't done much research yet), there is no RFC that specifies "ARP NUD". I have found that RFC1122 "2.3.2.1  ARP Cache Validation" does specify that Unicast polling of a neighbor using point-to-point ARPs can be use to check ARP cache entry validity. Also in passing I've noticed that Linux's implementation of ARP seems to use the states that it's IPv6 ND implementation uses, but I haven't yet verified that the behaviour is exactly the same. In other words, I think it is highly likely it can be applied to IPv4, but I wanted to see if "ARP NUD" was both permitted and also described somewhere. The follow on question is whether that description would have updated similarly to RFC4861.

> 3. I think a key eleemnt of the draft that should be made more clear in the 

> description is
>    that this mechanism is meant to be incrementally dpeloyable such that the 
> mayority
>    of devices on the (wireless) L2 domain do not have to be changed at all, and 
> that
>    the only expection against them is 2. (RFC6085 or equivalent for IPv4).
> 

I've realised through the discussion over the last few days that I need to separate out and make more clear some of these points. This first version was mainly to "throw it at the wall and see if it sticks."

>    To achieve this, the only requirement is to run IGMPv3/MLDv2, inhibit 
> MLDv1/IGMPv1v2,
>    and a device that wants to send multicast as unicast into the L2 domain needs 
> to
>    perform explicit tracking of the membership reports to know whom to send 
> unicast copies to. 
> 
>    And then there are optimizations such as RFC4861 which should be optional.

> 

The reason I though to use NUD to track the presence of the neighbors was that neighbor presence discovery is necessary to resolve the layer 2 information for the link-local MLDv2 source addresses, and NUD was tracking and verifying individual neighbors independently.

By default, MLDv2 general queries are only sent around every 2 minutes, and there isn't any reliability for that and the responses, meaning that if a query isn't received due to packet loss, or any of the reports are lost due to packet loss, the next opportunity to discover continued listener presence is in around 2 minutes time. So in the case of a single packet being lost, that might mean traffic is being sent to a non-existent listener for up to 4 minutes, which I think is getting a bit too close to the typical 5 minute aging of learned MAC addresses in switches. 

NUD will probe by default around every 30 seconds, and if it gets no response, will retransmit the probe around once a second. So it'd be much quicker and more reliable at detecting that the listener has disappeared than using MLDv2 General Queries. As MLDv2 General Queries are going to occur anyway, one optimisation that has occurred to me is that the corresponding reports from the listeners could be used as the evidence to Neighbor Discovery that the listener still exists, as a General Query and the corresponding reports are a two way transaction. The NUD probe could then be delayed for another 30 seconds. One drawback of this idea though would be that it would start roughly synchronising when NUD probes to all the listeners starts.

So I think using NUD (perhaps with MLDv2 GQ/Report hints) would be a better mechanism to use to track listener presence.

>    But having said all this, the fact that this can be done as a local 
> implementation
>    optimization in a device sending IP multicast into the L2 domain (whether 
> thats
>    a router, switch, host or whatever), it also means that this mechanism does 
> not
>    need to change any protocol signaling, and therefore the question is whether 
> this
>    can be standards track or just be informational.
> 

I'd think standards track due to the minor change required to RFC4861.

I'm not sure what happens with IPv4, as I haven't found an RFC describing "ARP NUD".

> 4. I don't particularily think that the operating models presented and the 
> first
>    sentence of the abstract are correct. Sending L2 (wireless) multicast has a 
> lot of
>    benefits, and it has a lot of downsides. It totally depends on what you try 
> to achieve.
> 

That's why I used "may".

Link layer multicasts should be used when they're better than what I've described. However, when link-layer multicasts are not going to perform as well as replication and link-layer unicasts, or when some of the other differences of replication and link-layer unicast are beneficial (i.e., better data confidentiality, better battery life on mobile devices like tablets, smartphones etc.), then what is described could be used.

The operating models were really just to highlight that not only could it be done on a per-group basis, but might also be useful to adopt for all groups on an interface, which may be a useful default model for an 802.11 interface.

>    It is clear that 802.11 is particularily challenged with native L2 multicast 
> because
>    they never defined a good resilience scheme as for unicast but so far not for 
> multicast.
>    Hopefully this will get fixed sometime. As long as thats not the case we 
> definitely
>    have to tke this into account as a constraint, but it would be good to 
> explain the
>    situation correctly in the doc.
> 

While I agree the doc is light on referenced evidence, I think it still describes the situation correctly, because I've done experiments of my own, and have seen bad multicast video over wifi, and have seen multicast traffic impact unicast performance. 

>    Wrt. the operational model, i could easily see that one wold start out with 
> unicasting,
>    but then change over to multicasting for a particular group/channel when the
>    total number of receivers tracked and bandwidth sent over the group/channel 
> is too large.

I can't think of a situation when falling back to multicast after using replicated link-layer unicast would be useful. I consider what I've described to be an alternative to link-layer multicast, to be used when either it is either known or likely from the outset that link-layer multicast will be inadequate, or when replication and link-layer unicast provides some useful advantages over link-layer multicast. Falling back to link-layer multicast wouldn't solve the replication and link-layer unicast performance problems if they've occurred (otherwise, use multicast in the first place), or would eliminate the useful advantages that replication and link-layer unicast provided.

I know it is possible to tune 802.11 to provide better multicast performance, however I understand that that tuning requires changing settings on all of the AP clients, as well as possibly on the AP, and if one new client which isn't tuned joins the link, performance drops back to before any tuning took place. This tuning would be beyond the capability of most residential home users.

Another thought I've had is that as wired switches are doing this type of packet replication and effectively link-layer unicasting out their individual ports, perhaps in the future the hardware doing this could also be adapted to increase the performance of replication and link-layer unicasting over 802.11.


>    You may also limit IP multicast to a particular bitrate, so if a specific 
> receiver
>    can not support that bitrate, it would need to get unicast, and then that 
> unicast
>    might need to be rate-limited so that the receiver downspeeds (assuming it is
>    using ALC which it really should). I am not aware if any product embodies 
>    this dyanmicac switching, i am just saying it sounds logical to me ;-)
> 

This generally sounds to me to be above layer 3 (and the interface to layer 2) and be in the domain of either things like RSVP, or applications/protocols that already provide reliable multicast, mainly because I think it is necessary for the receiver to indicate to the sender success or failure of packet delivery.

It could be done also be done on a link layer level, and I understand that 802.11 probably has enough mechanisms to be able to do this (e.g. acknowledgements and retransmissions). However, while 802.11 is the most likely use case for what I've described, I think there are benefits in keeping it generic, and applicable to any current or future link-layers which support both multicast and unicast transmission. For example, multicast over wired ethernet is (usually?) going to outperform replication and link-layer unicast (although, as a few people have pointed out off-list, switches are in effect doing replicating and link unicasting to emulate broadcast/multicast behaviour), however the other benefits may mean it is preferable to use what I've described instead of link-layer multicast.


>    So, given how many policies there could be to improve behavior of multicast 
> over
>    wireless, i think it would be more appropriate to make this draft 
> informational
>    and collect/document/discss these possible policies. Its not as if a specific 
> individual
>    policy seems to be ubiquoutously preferrable, and given how its all local 
> policy on the
>    device sending into the wireless L2, it doesn't seem to be the right 
> place for a
>    standard track doc. Going informational does hopefully not make a difference 
> for 
>    this doc being useful, because whether its standards track or not, customers 
> will hopefully 
>    put it into RFPs against their wireless equipment. Which also means its a 
> good idea
>    in these type of documents to give sticky names to individual functions such 
> as
>    different type of policies or optimizations, so customes in RFPs can refer to 
> those
>    sticky names (or section numbers).
> 
> 5. Not being really a protocol extension but logically device-local improvements 
> with
>    a bunch of policies possible makes it also a good candidate for MBoned as 
> opposed to a
>    protocol group (i think... ).
> 


Thanks again.

Best regards,
Mark.

> Cheers
>     Toerless
> 
>    tracking of 
>    that it is consisting purely of mechanisms that the accss-point type device 
> in a
>   
> 
> On Fri, Mar 29, 2013 at 03:18:06PM -0700, Mark Smith wrote:
>>  Hi Behcet,
>> 
>>  Thanks for your review and comments.
>> 
>> 
>> 
>>  >________________________________
>>  > From: Behcet Sarikaya <[e-mail]>
>>  >To: Dave Thaler <[e-mail]> 
>>  >Cc: Mark Smith <[e-mail]>; 
> "[e-mail]" <[e-mail]>; "[e-mail]" 
> <[e-mail]> 
>>  >Sent: Saturday, 30 March 2013 8:33 AM
>>  >Subject: Re: [MBONED] "MLDv2 Procedures for Link-Layer Unicast 
> Delivery of Multicast"
>>  > 
>>  >
>>  >Hi Mark,
>>  >
>>  >I read your draft. 
>>  >First of all I think you misunderstood RFC 6085 and based on a wrong 
> assumption you developed your solution.
>>  >
>>  >
>>  >What specifically do you think I've misunderstood? I was involved 
> in some of the conversations about that RFC while it was an ID, and all I think 
> it really is fundamentally saying is that an IPv6 packet with a multicast IPv6 
> destination address isn't required to have a link-layer multicast 
> destination address - it can be a link-layer unicast destination address if 
> useful or necessary.
>>  >
>>  >
>>  >
>>  >
>>  > I suggest you take a look at 
> http://tools.ietf.org/html/draft-sarikaya-netext-pmipv6-shared-link-01
>>  >on Netext for PMIPv6.
>>  >
>>  >
>>  >I'll have a read.
>>  >
>>  >I believe that we should use multicast delivery as much as possible on 
> the downlink if the link is a shared link.
>>  >
>>  >
>>  >
>>  >
>>  >I agree with using multicast as much as possible to reduce duplicate 
> packets sent over the network. The proposal here is to use packet replication 
> and link-layer unicasts when doing so would either exceed the performance of 
> link-layer multicast and/or overcome the negative effects of link-layer 
> multicast, such as when larger volumes of multicast traffic impact unicast 
> performance on the same link. Both of these issues occur on 802.11 links when 
> there volume of multicast traffic is in the order of a few megabits (or less, 
> I've been testing with around 2.5 Mbps of video)
>>  >
>>  >
>>  >The main use case I've been thinking about (which I'll make 
> more obvious in the next revision), is a multicast IPTV service scenario towards 
> a residential customer, where the residential customer has a 802.11 network in 
> their home rather than a wired one. IPv6 multicast/link-layer multicast would be 
> used all the way from the multicast source servers, across the service provider 
> network, up until the CPE in the residential customers' homes, gaining the 
> efficiencies of multicast. However, the CPE in the customer's house would 
> then use IPv6 multicast/link-layer unicast of the IPTV traffic over the 802.11 
> link to the customers end-hosts, saving them having to put in wired 
> infrastructure, or resorting to buying ethernet over power devices to make their 
> power cabling their wired infrastructure. Note that there is still a low level 
> of link-layer multicast traffic in the customers home, for RAs, ND, MLDv2 
> messages, DHCPv6, etc.. This proposal is not eliminating
>>   link-layer multicasts, just reducing them where possible and where useful.
>>  >
>>  >
>>  >That's not to say this proposal is only useful in an 802.11 
> scenario. It could be used in any scenario where packet replication and 
> link-layer unicasting would provide useful benefits over link-layer multicasts. 
> It generally increases the data confidentiality of the multicast IPv6 traffic, 
> and may help increase battery life of some devices.
>>  >
>>  >
>>  >Thanks very much,
>>  >Mark.
>>  >
>>  >
>>  >
>>  >Regards,
>>  >
>>  >Behcet
>>  >
>>  >
>>  >
>>  >
>>  >On Fri, Mar 29, 2013 at 3:55 PM, Dave Thaler 
> <[e-mail]> wrote:
>>  >
>>  >http://tools.ietf.org/html/draft-thaler-ngtrans-6to4-multicast
>>  >>is work we did back in 2000 on this same topic.   At the time, the 
> draft is written from
>>  >>the perspective of the 6to4 NBMA link, but the topic was discussed 
> (specifically by those
>>  >>in the acknowledgements section, and to a lesser extent by the 
> ngtrans WG as a whole)
>>  >>as being more generally applicable.
>>  >>
>>  >>There wasn't significant interest at the time and so the work 
> was dropped
>>  >>rather than updating the spec to use generic language.
>>  >>
>>  >>The same concept as in the above was however then used in 2001 in
>>  >>http://tools.ietf.org/html/draft-ietf-mboned-auto-multicast
>>  >>(still specific to a particular link type though), which after 11 
> years is now in the IESG :)
>>  >>
>>  >>The most relevant WG is MBoneD.
>>  >>
>>  >>-Dave
>>  >>
>>  >>
>>  >>> -----Original Message-----
>>  >>> From: [e-mail] [mailto:[e-mail]] On 
> Behalf Of
>>  >>> Mark Smith
>>  >>> Sent: Thursday, March 28, 2013 8:22 PM
>>  >>> To: [e-mail]
>>  >>> Subject: "MLDv2 Procedures for Link-Layer Unicast 
> Delivery of Multicast"
>>  >>>
>>  >>> Hi,
>>  >>>
>>  >>> The following is inspired by some work I did in around 
> 2010/2011 on a
>>  >>> multicast TV service, where residential customers with Wifi 
> networks
>>  >>> suffered from performance problems, and had to be advised to 
> buy
>>  >>> ethernet over power devices if they couldn't run wired 
> ethernet to the STB
>>  >>> attached to their TV.
>>  >>>
>>  >>> I've done some experiments on my home wifi network, using 
> VLC, and
>>  >>> switching between multicast and unicast IPv6 to test the 
> concept. The wifi
>>  >>> performance issues disappear when the video is delivered via 
> unicast UDP.
>>  >>> (If people want to play with VLC and IPv6 multicast, email me 
> off-list, as there
>>  >>> are a few issues e.g. the GUI doesn't accept some of the 
> IPv6 multicast
>>  >>> parameters that the command line does).
>>  >>>
>>  >>> I think there is a possibility the technique could also be 
> applied to
>>  >>> IGMPv3/ARP, although it depends on whether ARP has been 
> implemented
>>  >>> in a similar manner to IPv6 ND (e.g. an ARP equivalent to 
> NUD). My
>>  >>> understanding is that Linux has implemented ARP this way. 
> I'll do some more
>>  >>> research to verify this.
>>  >>>
>>  >>> I though I'd post it to see if there is merit in the idea 
> and it is worth spending
>>  >>> more of my time on. I looked for a IETF group more suitable to 
> this, but didn't
>>  >>> seem to be able to find one. If there is one, please let me 
> know.
>>  >>>
>>  >>> Review and comments most appreciated.
>>  >>>
>>  >>> Thanks very much,
>>  >>> Mark.
>>  >>>
>>  >>>
>>  >>> ----- Forwarded Message -----
>>  >>> > From: "[e-mail]" 
> <[e-mail]>
>>  >>> > To: [e-mail]
>>  >>> > Cc:
>>  >>> > Sent: Friday, 29 March 2013 1:45 PM
>>  >>> > Subject: New Version Notification for
>>  >>> > draft-smith-mldv2-link-unicast-00.txt
>>  >>> >
>>  >>> >
>>  >>> > A new version of I-D, 
> draft-smith-mldv2-link-unicast-00.txt
>>  >>> > has been successfully submitted by Mark Smith and posted 
> to the IETF
>>  >>> > repository.
>>  >>> >
>>  >>> > Filename:     draft-smith-mldv2-link-unicast
>>  >>> > Revision:     00
>>  >>> > Title:         MLDv2 Procedures for Link-Layer Unicast 
> Delivery of
>>  >>> > Multicast
>>  >>> > IPv6
>>  >>> > Creation date:     2013-03-29
>>  >>> > Group:         Individual Submission
>>  >>> > Number of pages: 7
>>  >>> > URL:
>>  >>> > 
> http://www.ietf.org/internet-drafts/draft-smith-mldv2-link-unicast-00.
>>  >>> > txt
>>  >>> > Status:
>>  >>> > 
> http://datatracker.ietf.org/doc/draft-smith-mldv2-link-unicast
>>  >>> > Htmlized:
>>  >>> > 
> http://tools.ietf.org/html/draft-smith-mldv2-link-unicast-00
>>  >>> >
>>  >>> >
>>  >>> > Abstract:
>>  >>> >    Some multi-access link-layer technologies typically 
> not provide
>>  >>> > good
>>  >>> >    IPv6 multicast performance, using link-layer 
> multicasts, when the
>>  >>> >    volume of multicast traffic is significant.  It would 
> be possible
>>  >>> > to
>>  >>> >    replicate and then link-layer unicast multicast IPv6 
> traffic to
>>  >>> >    interested listeners to overcome these link-layer 
> performance
>>  >>> >    limitations.  This memo describes MLDv2 and IPv6 
> neighbor discovery
>>  >>> >    procedures to support link-layer unicast delivery of 
> multicast IPv6
>>  >>> >    traffic.
>>  >>> >
>>  >>> >
>>  >>> >
>>  >>> >
>>  >>> >
>>  >>> > The IETF Secretariat
>>  >>> >
>>  >>> 
> --------------------------------------------------------------------
>>  >>> IETF IPv6 working group mailing list
>>  >>> [e-mail]
>>  >>> Administrative Requests: 
> https://www.ietf.org/mailman/listinfo/ipv6
>>  >>> 
> --------------------------------------------------------------------
>>  >>_______________________________________________
>>  >>MBONED mailing list
>>  >>[e-mail]
>>  >>https://www.ietf.org/mailman/listinfo/mboned
>>  >>
>>  >
>>  >
>>  >
>>  _______________________________________________
>>  MBONED mailing list
>>  [e-mail]
>>  https://www.ietf.org/mailman/listinfo/mboned
> 
> -- 
> ---
> Toerless Eckert, [e-mail]
> Cisco NSSTG Systems & Technology Architecture
> SDN: Let me play with the network, mommy!
>