Re: [Idr] Use of BGP for Opaque Signaling

[Gunter Van De Velde <[e-mail]>]

One of the questions in the equation is if signalling of Opaque information needs potentially modified/different path selection mechanisms compared to a placeholder IPv6 NLRI + wide community.

Ciao,
G/



> On 13 Mar 2016, at 20:12, Robert Raszuk <[e-mail]> wrote:
> 
> Petr,
> 
> > E.g., you need an NLRI of some kind to attach the communities, i.e. every originator 
> > needs to inject a prefix - which isn't too bad implementation wise, but ties information 
> > distribution to reachability info
> 
> True about need for NLRI. Not so true about putting equal sign between NLRI and reachability in existing SAFIs IMHO. 
> 
> Let's consider that we define a special (v6) prefix which will be used as an NLRI prefix for your opaque distribution. Part of such prefix (suffix) can be your key (verbatim) ... values on the other hand can be carried in wide communities or if you insist in your new attribute. We could even have one for v4 too - say 240.*.*.*. 
> 
> It will be much easier to filter or accept such UPDATE messages and policy required will be pretty much the same as filtering to accept or drop some keys from your new SAFI. Besides AFI/SAFI 2/1 is already there both on RRs or ASBRs hence interested parties may not need to upgrade OS running on those routers. 
> 
> Besides we have already seen from the other drafts (for example Stephane's https://tools.ietf.org/html/draft-litkowski-idr-bgp-timestamp-02 <https://tools.ietf.org/html/draft-litkowski-idr-bgp-timestamp-02>) the need to define such special NLRIs to carry opaque information in existing SAFIs which do not really need to be processed by BGP implementations. 
> 
> > Furthermore, wide communities could be interpreted by BGP process
> 
> Only if you configure them to be processed or they are registered as public and their types are maintained by IANA. The default is opaque just as you need it. 
> 
> Thoughts ? 
> 
> Best,
> Robert.
> 
> 
> On Sun, Mar 13, 2016 at 12:37 PM, Petr Lapukhov <[e-mail] <mailto:[e-mail]>> wrote:
> Robert, I totally looked at the wide communities before trying the opaque stuff. However, there was some semantical awkwardness in regard to transporting opaque data. E.g., you need an NLRI of some kind to attach the communities, i.e. every originator needs to inject a prefix - which isn't too bad implementation wise, but ties information distribution to reachability info, requiring a policy to ignore this reachability information. I really don't like adding more policies :) 
> 
> Furthermore, wide communities could be interpreted by BGP process, and hence their implementation is a bit more involved and requires more interop testing. The opaque AFI simply acts as a pipe to disseminate bits in the domain (e.g. we use thrift to encode values). Implementing it requires accepting the new AFI/SAFI and skipping IP-specific policy code (e.g. filtering only on attributes).
> 
> Most importantly, opaque data is "free" of any structure that could require further standardization :) This is on purpose - I was hoping to have this channel defined once, and let it go without any further need of defining additional sub-structures...
> 
> PS
> The 64K messages would be pretty awesome! :) Having the 4K limit in some cases requires annoying value fragmentation. Ugh.
> 
> From: [e-mail] <mailto:[e-mail]> [[e-mail] <mailto:[e-mail]>] on behalf of Robert Raszuk [[e-mail] <mailto:[e-mail]>]
> Sent: Sunday, March 13, 2016 7:59 AM
> To: Gunter Van De Velde
> Cc: Petr Lapukhov; idr wg
> Subject: Re: [Idr] Use of BGP for Opaque Signaling
> 
> Hi Gunter,
> 
> For dissemination of information produced by network applications we already have today various tools be it ansible, chef, puppet etc ... 
> 
> If there is need to use inter-domain distribution of such applications I see no reason why not to define registered or private wide bgp community and use this as transport instead of new SAFI on already established BGP sessions. Note that wide BGP community precisely like new attribute being proposed here can carry opaque to BGP data. 
> 
> Hi Petr,
> 
> Do you see any issue with the above wide community approach to address you needs ? Especially for bootstrapping the transport it is even easier as you do not need to enable new SAFI. And those are already IDR WG docs. 
> 
> https://tools.ietf.org/html/draft-ietf-idr-wide-bgp-communities-01 <https://urldefense.proofpoint.com/v2/url?u=https-3A__tools.ietf.org_html_draft-2Dietf-2Didr-2Dwide-2Dbgp-2Dcommunities-2D01&d=CwMFaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=LU_vJaM_EQu1Ssm35j2xlA&m=h-7EJnudbBhYxtlkgN-DvH3EQwxZkBfN9YlkWGqrg8g&s=cGH3bKxTAm1Xiy1qkwo3f69UjqvwTTjKqu2W4nrYhG8&e=>
> https://tools.ietf.org/html/draft-ietf-idr-registered-wide-bgp-communities-01 <https://urldefense.proofpoint.com/v2/url?u=https-3A__tools.ietf.org_html_draft-2Dietf-2Didr-2Dregistered-2Dwide-2Dbgp-2Dcommunities-2D01&d=CwMFaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=LU_vJaM_EQu1Ssm35j2xlA&m=h-7EJnudbBhYxtlkgN-DvH3EQwxZkBfN9YlkWGqrg8g&s=bApGAS0Vc65I-5JXPEbRMymvicovHKZYArXUwbZcRaU&e=>
> 
> As far as limit of 4K let's all note that we already have a draft addressing this and moving the limit to 65K: https://tools.ietf.org/html/draft-ietf-idr-bgp-extended-messages-11 <https://urldefense.proofpoint.com/v2/url?u=https-3A__tools.ietf.org_html_draft-2Dietf-2Didr-2Dbgp-2Dextended-2Dmessages-2D11&d=CwMFaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=LU_vJaM_EQu1Ssm35j2xlA&m=h-7EJnudbBhYxtlkgN-DvH3EQwxZkBfN9YlkWGqrg8g&s=44sn2G4s0yMfrbNf7GGtcfd1InmvpRE6axtKhrM3bVU&e=>
> 
> Cheers,
> r.
> 
> 
> 
> 
> 
> On Sun, Mar 13, 2016 at 5:11 AM, Gunter Van De Velde <[e-mail] <mailto:[e-mail]>> wrote:
> imo the promise of a proposal like this one, is that it allows Application development at DevOps speed and removes slowness of industry bodies for innovation.
> 
> I believe it is a valid problem space that is being addressed. Maybe some features/optimisations from EVN could be transplanted? (i.e. Designated router selection, aliasing and mass-withdraw)
> 
> G/
> 
> 
>> On 13 Mar 2016, at 04:20, Petr Lapukhov <[e-mail] <mailto:[e-mail]>> wrote:
>> 
>> > In fact I am not convinced p2mp is the best way to go with here. 
>> > Pub-sub without RTC hacks may be much better - don't you honestly think ? 
>> 
>> Not exactly, but maybe I misunderstood what p2mp means :) The main advantage of using BGP (at least in our environment) comes from the fact that it's already bootstrapped when on-switch (or on-server) applications start coming up. And it stays there even when the servers hosting external key-value store may be gone. 
>> 
>> Aside from this, I fully agree - it does not offer much more than any key-value store. And it has its own constraints, like ~4K value size. Plus, it's not even a proper pub-sub, since there is no discreet message semantic - only state dissemination :)
>> 
>> From: [e-mail] <mailto:[e-mail]> [[e-mail] <mailto:[e-mail]>] on behalf of Robert Raszuk [[e-mail] <mailto:[e-mail]>]
>> Sent: Saturday, March 12, 2016 5:25 PM
>> To: Petr Lapukhov
>> Cc: [e-mail] <mailto:[e-mail]>; idr wg
>> Subject: Re: Use of BGP for Opaque Signaling
>> 
>> ​Hello Petr,​
>>  
>> Re: Impact on routing. I should be more explicit stating that the idea is for BGP to not interpret the key-value mappings in any way.
>> 
>> ​That was very clear already. But if you look at number of BGP implementations recognizing the AFI/SAFI coming over single session is not done immediately in the BGP message parsing. That is my main point suggesting transport instance port. 
>> 
>> I no longer believe multisessions will do it neither that we could off load it as implementation specific thing.​
>> 
>>  
>> The consumer (running on same device, possibly) may still influence routing based on its interpretation of data, either by accessing the platform SDK or by performing BGP routing injection over the corresponding AFI/SAFI.
>> 
>> ​That is exactly the issue. The only way BGP-LS went by was the claim about their explicit disjointed topology from existing BGP. It was well positioned that BGP-LS can be completely incongruent with "routing" BGP. ​
>> 
>> ​The possible channels the IGP information may impact routing were classified as out of the band likely coming via some other ways from controllers (I2RS, APIs, routing BGP SAFI etc ...). 
>> 
>>  
>> Re: Transport instance. I think it's a great idea to run this on separate instance, but right now we typically run just one process, and it adds some operational overhead to maintain two of them. Also, often this extension could be used just with out-of-band reflection. This is why I feel it is rather better to keep this applicable to any type of BGP instance.
>> 
>> 
>> ​Imagine that you will not get far in this BGP proposal. Then what is left is nothing else like a new process talking over one of few message buses today to distribute the set of information you need. 
>> 
>> In fact I am not convinced p2mp is the best way to go with here. Pub-sub without RTC hacks may be much better - don't you honestly think ? 
>> 
>> Best,
>> R.​
>> 
>>  
>> Re: Scope of application. The main goal was to use this in data-center, but I can theorize this could be used among few operators that are willing to develop on-device applications and exchange data using the pre-established channels (e.g. the WISER use case). In this situation, limiting scope to just private ASNs could be tricky. Furthermore, non-private ASNs could be used in data-center as well (just not exposed outside...).
>> 
>> Re: Transitive attribute. Ideally, it would be great to carry the value in next-hop field of MP_REACH_NLRI, but the length is limited to 8 bytes, hence the new attribute. I do not have strong arguments to making this transitive, it should be fine to go with non-transitive and corresponding error handling. 
>> 
>> Re: Security. Since we do not directly influence path selection with new data, the threats could be limited to those targeting BGPinfra itself e.g. flooding with information, disrupting session etc. There could be questions about information authenticity, but this could be left up to the applications to decide (e.g. signing the key-value bindings).
>> 
>> Regards,
>> 
>> Petr
>> 
>> From: [e-mail] <mailto:[e-mail]> [[e-mail] <mailto:[e-mail]>] on behalf of Robert Raszuk [[e-mail] <mailto:[e-mail]>]
>> Sent: Saturday, March 12, 2016 1:10 PM
>> To: Petr Lapukhov; [e-mail] <mailto:[e-mail]>
>> Cc: idr wg
>> Subject: Use of BGP for Opaque Signaling
>> 
>> ​Petr, Eben,
>> 
>> I have few questions and comments to your proposal. 
>> 
>> - The draft ​attempts to reuse BGP mechanism for p2mp distribution of opaque information. That alone is not a bad thing :) However when you make reference to BGP-LS it needs to be observed that BGP-LS is not position to directly impact routing (ie content of RIB and FIBs) unlike one of the use cases described in your draft Intro paragraph. That means that unlike BGP-LS it must co-exist with routing BGP on the same BGP speakers. 
>> 
>> Here I would like to bring back my proposal to define new port for Transport Instance BGP https://tools.ietf.org/html/draft-raszuk-ti-bgp-01 <https://urldefense.proofpoint.com/v2/url?u=https-3A__tools.ietf.org_html_draft-2Draszuk-2Dti-2Dbgp-2D01&d=CwMFaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=LU_vJaM_EQu1Ssm35j2xlA&m=75Q-nvX2RhhxYm-r2BIxFhtIpKzT9pEgGx9UWjq6fT4&s=kK71eSqWBITalG1Y1iNWaiHaSA7vrMMnL4lkieaq2uY&e=> which at the time of discussion was supposed to be shortly addressed by dynamic port allocations of multisessions which to the best of my knowledge never happened. 
>> 
>> So first comment is that if we go with reusing BGP-4 for this type of signalling let's enforce it's early separation from "classic" BGP. 
>> 
>> - Second comment is related to the scope of use. If the scope is private - let's perhaps enforce it more explicitly such that by the spec it is illegal AFI/SAFI pair in the public Internet. For example by stating that it is illegal to have non private AS in the AS_PATH of such UPDATE msg. 
>> 
>> - Third the draft states that there is no new security implications .. well if the scope is non private I am afraid that is no the case. As you have seen (and still seeing) securing BGP UPDATE MSG is MP_REACH content dependent. If content is free form I am afraid it will be rather hard to secure it well :) 
>> 
>> - Last as opaque value container you are defining an transitive attribute ? Is the intention to use this attribute in existing AFI/SAFIs as well ? Otherwise if not why do you need this to be transitive if capability negotiation already may implicitly define it's support by new AFI/SAFI.
>> 
>> Cheers,
>> R.
>> 
>> PS. 
>> 
>> > There is an ongoing discussion around avoiding NLRI (key) collisions.
>> 
>> Where is such discussion taking place ? Shouldn't IDR be perhaps in the loop a bit :) 
>> 
>> 
>> 
>>  
>> 
>> _______________________________________________
>> Idr mailing list
>> [e-mail] <mailto:[e-mail]>
>> https://www.ietf.org/mailman/listinfo/idr <https://urldefense.proofpoint.com/v2/url?u=https-3A__www.ietf.org_mailman_listinfo_idr&d=CwMFaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=LU_vJaM_EQu1Ssm35j2xlA&m=h-7EJnudbBhYxtlkgN-DvH3EQwxZkBfN9YlkWGqrg8g&s=UEjDtCq5z4uORK2BA9Wqg6I8MnbxxtajRr33DW-dWWM&e=>
> 
> 
> _______________________________________________
> Idr mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/idr