2001 Message-Id: Date: Mon, 31 Dec 2001 14:47:42 -0800 From: "Putzolu, David" Subject: Applicability Statement Draft: Ready to be WG Document? MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" All, At the ForCES meeting during the last IETF (Salt Lake City, Monday December 10), Mark Handley presented an applicability statement that had been uploaded a month prior. URL: http://search.ietf.org/internet-drafts/draft-crouch-forces-applicability-00. txt The consensus of those present in the room was that this document is ready to be accepted as a working group document. I would like to verify this with the list. Please comment on the following: * Is this document ready to be accepted as a WG document? Note that acceptance does not finalize the document, it merely associates it with the working group. * If not, what changes need to be done in order to make this acceptable as a WG document? Please send your comments to the list. Thanks, David Putzolu (co-chair) 2001 Message-Id: Date: Mon, 31 Dec 2001 14:20:30 -0800 From: "Putzolu, David" Subject: ForCES Meeting Minutes, IETF 52 MIME-Version: 1.0 Content-Type: multipart/mixed; boundary="----_=_NextPart_000_01C19249.5B548C20" This message is in MIME format. Since your mail reader does not understand this format, some or all of this message may not be legible. ------_=_NextPart_000_01C19249.5B548C20 Content-Type: text/plain; charset="iso-8859-1" All, Please find attached the first draft of the meeting minutes from the ForCES meeting at IETF 52. Please send any comments or corrections that need to be included. -David and Patrick ------_=_NextPart_000_01C19249.5B548C20 Content-Type: text/plain; name="ForCESMinutes.txt" Content-Transfer-Encoding: quoted-printable Content-Disposition: attachment; filename="ForCESMinutes.txt" ForCES Meeting Minutes =96 Dec 10, 2001, 1930-2200 Notetakers: David Putzolu, Patrick Droz draft-ietf-forces-requirements-01 =96 Avri Doria Focus on changes since last draft =96 open issues that were = unresolved=20 in -00 Changes: Tightened up definitions Pre-association phase and post-association phase =96 what is=20 pre-association phase vs. post-association phase? Agreed that=20 pre-association was discovery (who is my FE/who is my CE?) and=20 post-association was connection establishment and control of FE=20 by CE Multiple-CE issue =96 decided to allow multiple CEs but declare=20 that CE-CE communication would be out of scope for the ForCES=20 protocol =96 re-charter later if we want to do it FE Model =96 must express logical packet processing capabilities=20 of the FE Two layer model Management issue =96 NE appears as one entity on the network =96=20 including management.=20 However, people may still want to use a MIB for their FE, etc.=20 Recognized and accepted that external management may access a FE=20 for monitoring and transitioning from pre- to post-association = phase. Transport Protocol =96 for IP networks an RFC2914 compliant = transport=20 will be used. Still need application layer reliability (e.g.=20 acknowledge command success) Questions? Kathleen Nichols =96 PacketDesign: Question on QoS section =96 = reference=20 to a draft =96 should correct references to be to RFCs not drafts V. Sharma =96 Q: When you have mgmt traffic =96 typically you go to = the=20 CE directly rather than the FE. A: We assume that most access is = through the CE =96 just accepting the fact that some may want to = go=20 directly to FE =96 cannot preclude it. David McDysan - Worldcom =96 Q: Messages received by FE that need = to be=20 processed by CE. May want to mention in requirements that ability = to place controllers or multiple controllers in the network is=20 needed. Be more detailed Ram Gopal =96 Nokia =96 Q: Why does the FE Manager talk to the FE = =96 should=20 it talk via the CE? A: Pre-association phase is where FE mgr. = needs to talk to FE =96 it is the entity that=20 informs the FE who its CE is. Q: Is CE and FE mgr in scope of = ForCES?=20 A: Yes, CE and FE managers are outside scope of protocol, but = exist.=20 Comment: Should elucidate that these are not part of the protocol = or=20 say that the minimum requirements and say rest is off the table=20 (e.g. we don't care).=20 ? - Q: Maybe in this document we need to address the issue of = crashing of=20 FE or CE. A: Will go back and check if there, if not will add it. ? - Comment: May often have management PDUs going through the FE to = the=20 CE =96 is an engineering issue Alan DeKok =96 Solidum: Q: 6.2.1.1 Some short text on = classification =96=20 DiffServ model draft should be referenced here. Chairs: Last call? No strong opinion in the room =96 decided to do = another=20 iteration of the draft before going to last call. draft-crouch-forces-applicability-00 =96 ForCES Applicability Statement = =96 Mark=20 Handley Why does ForCES need one? Try to make sure we are all on the same=20 page =96 say both what we want to do and what ForCES isn't. Don't expect this to become standard until ForCES protocol itself = goes=20 to RFC ? - Comment: This is perfect for an informational RFC =96 good place = to=20 discuss important issues like what uses of ForCES. A: Don't want to = micromanage but do want to give good guidelines. Manav Mishra - Intel: FE to FE communication question =96 haven't = mentioned=20 FE to FE or CE to CE. A: Still don't have consensus. Margaret Wasserman =96 Wind River: Value of writing this now? Isn't = this=20 requirements? A: These are more bounding the problem than = specifying=20 requirements for a solution.=20 ? - Comment: We are scoping what the design and application of the = protocol=20 should look at =96 don't use it for Dave McDysan - WorldCom: Requirements should take inspiration from=20 this =96 maybe should defer this till later. Also, label-switching = being=20 out of scope may make this work more experimental as we may need = that=20 function to be used in production. A: IESG initially directed us to = focus on the things that don't exist =96 IP. Bill Fenner =96 Routing AD: Look at applicability statement as = something=20 saying what we don't have to solve as opposed what we do have to = solve.=20 Thus lets us focus better on what we do need to solve. A: Agree = completely.=20 Reason this document was people were talking about using ForCES for = thing=20 such as controlling an entire WAN from one central box, etc., - = wanted to=20 focus ForCES more to not solve all possible configurations. Chairs: Should we make this a working group document? Rough = consensus in=20 the room for making this a WG document. Will take to list for = confirmation. draft-anderson-forces-model-00 =96 ForCES Model and Architecture =96 = Todd=20 Anderson Architecture Picture ? - Comment: Can be extremely complicated to have multiple CE=20 coordination =96 would be interesting to see on the list who = wants to=20 do this. Ram Gopal: What if one CE set controls one FE, another set con Margaret Wasserman: CE sets useful for failover, etc.=20 Ram Gopal - Any CE should be able to talk to any FE. ? =96 From ForCES protocol perspective should we have one CE or = multiple=20 CEs? Is it relevant? A: Need to just explain this model in order = to have=20 the FE be aware of multiple CEs. Comment: Don't want to duplicate = what=20 is in other working groups. A: GR is out of scope =96 most seem = to think=20 that=20 Avri Doria: Comments: FEs are logical FEs; Many CEs can be talking = to same=20 FE at same time =96 up to CEs to keep themselves coordinated. CEs = also=20 need a way to discover other CEs =96 that may need to be dealt = with. A:=20 Maybe that is part of GC. Margaret Wasserman: Do the FEs need to tell CEs if other CEs have = done=20 something? If we can keep that out of scope things are simpler. Avri Doria: Agreement =96 FEs should not need to keep CEs = coordinated. Ram Gopal: This brings out some open issues in requirements and=20 applicability documents. FE Model ? =96 When do you assign namespace =96 part of capability exchange? = A: IANA=20 assign. Q: What if two FEs have same capabilities? A: Send same = info to=20 CE =96 use the same namespace of FE types. Q: How does CE = differentiate=20 FEs? A: Via protocol connection ID (endpoint ID). Comment: FE has = to=20 know which namespace it belongs to. A: FE has to know values of = names=20 for its capabilities. Comment: How is range conveyed prior to = protocols=20 like RIP, etc. running? A: Range is used to say what acceptable = values=20 a FE can support for the value of some configured protocol. Mark Handley =96 Where do stage instances appear from? Some are = inherent=20 abilities of FE, some might be configured from CE =96 i.e. = configure together=20 these five things. A: FE exposes a fixed topology of stages. = Comment:=20 Should not rule out reconfiguration at the start. Manav Mishra =96 Intel =96 Q: Do stages or capabilities and stages = get reported=20 to CE? A: Both =96 first list supported stages, then their types, = then the=20 topology of their graft. Q: What about how things are restricted? = A:=20 Reported on a per-property basis. V. Sharma =96 Q: If you do reconfiguration, is traffic interrupted? = A: Probably Consensus that due to the newness of this draft it should go through = another=20 iteration before acceptance as a WG document. General Comments Bill Fenner (AD) said that the WG has documents we want to get = feedback soon.=20 Open issues should be posted to the mailing list that people can = comment on=20 the unsolved or unresolved problems.=20 There were also a number of comments made to include FE-to-FE = communications. ------_=_NextPart_000_01C19249.5B548C20-- 2001 Message-Id: Date: Tue, 18 Dec 2001 09:43:17 -0800 From: "Putzolu, David" Subject: ForCES Meeting Summary, IETF 52 Comments: To: "Bill Fenner (E-mail)" , "Randy Bush (E-mail)" MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" The ForCES working group met at IETF 52 and reviewed several key documents. Changes to the WG requirements document were presented and accepted, however, after discussion it was decided that one more rev was needed before last call. An individual contribution of an applicability statement was presented and accepted as a working group draft. An individual contribution of a framework/model was also presented and met with strong interest but was not accepted as a WG document due to its newness and lack of time for proper review. At the end of the meeting some discussion was held about WG scope. CE to CE communication (e.g. how to perform distributed BGP RIB/FIB computation) was mentioned as out of scope, however, designing the CE/FE protocol with considerations so that such distributed computations would be feasible was accepted as a possibility. Also discussed was the possibility of examining FE to FE communications. 2001 Message-Id: Date: Thu, 13 Dec 2001 15:29:54 -0600 From: Ram Dantu Subject: Re: ForCES scope MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable I could not be there in WG meeting due=20 to schedule conflict with Midcom. I agree with this opinion. I think items like "BGP in distributed environment"=20 may be more of CE<-->CE and we may not be able to handle this in this WG. Best Regards Ram Dantu -----Original Message----- From: Patrick Droz [mailto:dro@zurich.ibm.com] Sent: Monday, December 10, 2001 9:38 AM To: FORCES@PEACH.EASE.LSOFT.COM Subject: ForCES scope I think there are 2 issues I would like to bring up here. It might be interesting to talk about the building of a distributed routing computer but I think it does not belong in this WG. ForCES should be completely independent of a particular routing protocol. In addition, I am the opinion that ForCES should concentrate on the CE <-> FE communication before talking about CE <-> CE stuff. Regards, Patrick -- Dr. Patrick Droz | dro@zurich.ibm.com IBM Zurich Research Laboratory | http://www.zurich.ibm.com/~dro Saumerstrasse 4 | Tel. +41-1-724-85-25 CH-8803 Rueschlikon/Switzerland | Fax. +41-1-724-85-78 2001 Message-Id: Date: Tue, 11 Dec 2001 09:49:53 -0800 From: "Anderson, Todd A" Subject: Re: comments on architecture draft Comments: To: "hadi@znyx.com" MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Let me assure the group that the model draft is completely unimcumbered by any IP issues as far as I know. The draft is not based on anything in the NPF. It's biggest inspiration is probably the DiffServ informal management model. Todd -----Original Message----- From: Jamal Hadi Salim [mailto:hadi@znyx.com] Sent: Monday, December 10, 2001 9:52 PM To: FORCES@PEACH.EASE.LSOFT.COM Cc: todd.a.anderson@intel.com Subject: comments on architecture draft I am a little concerned in general about the draft direction. I have a lot of comments but i might post them later or maybe whip an altrenative draft; Granted this is version 0 but i think what bothers me is that it comes out as a design document of some form; what is this based on? is this based on what the NPF has in place? If it is, why is the NPF not releasing their document(s) publicly? Are there any patents if it is derived from NPF? etc etc cheers, jamal 2001 Message-Id: Date: Tue, 11 Dec 2001 02:50:16 -0500 From: Jamal Hadi Salim Organization: Znyx Networks Subject: applicability draft Comments: cc: alan.crouch@intel.com, mjh@icsi.berkeley.edu MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: base64 DQpGaW5hbGx5IHJlYWQgdGhlIGRyYWZ0OyBDb21tZW50czoNCg0KLSBUaGUgZHJhZnQgdHJpZXMg dG8gZ2V0IGF3YXkgZnJvbSBzcGVjaWZpY3MgKGJ5IHVzaW5nIHRoZSBzdGQgZGlzY2xhaW1lciAN CiJmb3IgZXhhbXBsZSIpIHdoaWNoIGlzIGEgZ29vZCB0aGluZzsgaG93ZXZlciBzb21lIG9mIHRo ZSBjb21tZW50cyBtaWdodCBiZSANCmRlZW1lZCB0byBzdWdnZXN0IGEgY2VydGFpbiBhcmNoaXRl Y3R1cmUuIFRvIGdpdmUgb25lIGV4YW1wbGU6IHN1Z2dlc3RpbmcgdGhlIA0KdXNlIG9mIEZvckNF UyBwcm90b2NvbCBmb3IgQ0VzIHRvIEZFcyBmb3IgY2FwYWJpbGl0eSBleGNoYW5nZS4gSSB3b3Vs ZCBzYXkgDQp0aGF0IGZpdHMgaW4gdGhlIGRpc2NvdmVyeSBwaGFzZTsgYW5kIGV2ZW4gdGhlIGRp c2NvdmVyeSBwaGFzZSBtaWdodCBub3QgDQpuZWNlc3NhcmlseSBoYXZlIGEgY2FwYWJpbGl0eSBl eGNoYW5nZSBwaGFzZSAoaSBkb250IHdhbm5hIGdvIGludG8gZGV0YWlscykuDQoNCi0gSSBjYW4g c2VlIHRoZSBwb2xpdGljYWwgZGlzY2xhaW1lcnMgb2Ygc2VjdGlvbiAzLjEuMTAgd2l0aCByZWdh cmRzIHRvIA0KbWlkY29tIGFuZCBzZWN0aW9ucyA0LjEgd2l0aCByZWdhcmRzIHRvIEdTTVAgYW5k IG1lZ2Fjby4gSSBkb250IHVuZGVyc3RhbmQgDQp3aHkgdGhlcmUgYXJlIHRlY2huaWNhbCBsaW1p dGF0aW9ucyB0byBlbmZvcmNlIHRoZSAib3V0IG9mIHNjb3BlIHN0YXRlbWVudCIuIA0KV2h5IGRv IHlvdSB3YW50IHRvIGxpbWl0IGhvdyBGb3JDRVMgZ2V0cyB1c2VkPw0KDQotIDMuMS41OiBmaXJz dCBzZW50ZW5jZSAiLi4uIHRvIENFcyIgdGhhdCBzaG91bGQgcmVhZCBhcyAiLi4uIGZyb20gQ0Vz Ig0KDQozLjMuMjoNCi0gZGlkbnQgcXVpZXQgdW5kZXJzdGFuZCBzZWNvbmQgcGFyYWdyYXBoLiBX aHkgd291bGQgaXQgYmUgYSBwcm9ibGVtIGZvciB0aGUgDQpGRTwtPkNFIGxpbmsgZ29pbmcgZG93 bj8gaWYgdGhlcmUgd2FzIGluZm9ybWF0aW9uIGFscmVhZHkgY2FjaGVkIG9uIHRoZSBGRSwgDQp3 aHkgaXMgdGhpcyB1bnVzYWJsZT8NCg0KNC4yIGJ1bGxldCAxOiBIb3cgYXJlIHlvdSBkZXJpdmlu ZyB0aGUgdXBwZXIgYm91bmQgb2YgMyBob3BzPw0KNC4yIGJ1bGxldCAyOiB3aHkgdGhlIGxpbWl0 YXRpb24gb2Ygbm9uIGR5bmFtaWMgcm91dGluZz8gSSB0aGluayB5b3UgDQpleHBsYWluZWQgYnV0 IGkgbmV2ZXIgcXVpZXQgZ290IGl0Lg0KDQpjaGVlcnMsDQpqYW1hbCAgICAgICAgICAgICAgICAg IA0KDQo= 2001 Message-Id: Date: Tue, 11 Dec 2001 00:51:33 -0500 From: Jamal Hadi Salim Organization: Znyx Networks Subject: comments on architecture draft Comments: cc: todd.a.anderson@intel.com MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: base64 DQoNCkkgYW0gYSBsaXR0bGUgY29uY2VybmVkIGluIGdlbmVyYWwgYWJvdXQgdGhlIGRyYWZ0IGRp cmVjdGlvbi4gSSBoYXZlIGEgbG90IG9mIA0KY29tbWVudHMgYnV0IGkgbWlnaHQgcG9zdCB0aGVt IGxhdGVyIG9yIG1heWJlIHdoaXAgYW4gYWx0cmVuYXRpdmUgZHJhZnQ7IA0KR3JhbnRlZCB0aGlz IGlzIHZlcnNpb24gMCBidXQgaSB0aGluayB3aGF0IGJvdGhlcnMgbWUgaXMgdGhhdCBpdCBjb21l cyBvdXQgYXMgDQphIGRlc2lnbiBkb2N1bWVudCBvZiBzb21lIGZvcm07IHdoYXQgaXMgdGhpcyBi YXNlZCBvbj8gaXMgdGhpcyBiYXNlZCBvbiB3aGF0IA0KdGhlIE5QRiBoYXMgaW4gcGxhY2U/IElm IGl0IGlzLCB3aHkgaXMgdGhlIE5QRiBub3QgcmVsZWFzaW5nIHRoZWlyIA0KZG9jdW1lbnQocykg cHVibGljbHk/IEFyZSB0aGVyZSBhbnkgcGF0ZW50cyBpZiBpdCBpcyBkZXJpdmVkIGZyb20gTlBG PyBldGMgZXRjDQoNCmNoZWVycywNCmphbWFsDQo= 2001 Message-Id: Date: Mon, 10 Dec 2001 16:38:22 +0100 From: Patrick Droz Organization: IBM Research Division Subject: ForCES scope Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit I think there are 2 issues I would like to bring up here. It might be interesting to talk about the building of a distributed routing computer but I think it does not belong in this WG. ForCES should be completely independent of a particular routing protocol. In addition, I am the opinion that ForCES should concentrate on the CE <-> FE communication before talking about CE <-> CE stuff. Regards, Patrick -- Dr. Patrick Droz | dro@zurich.ibm.com IBM Zurich Research Laboratory | http://www.zurich.ibm.com/~dro Saumerstrasse 4 | Tel. +41-1-724-85-25 CH-8803 Rueschlikon/Switzerland | Fax. +41-1-724-85-78 2001 Message-Id: Date: Sat, 8 Dec 2001 21:16:43 +0530 From: Mareline Sheldon Subject: Re: BGP in a distributed environment Comments: cc: ram.gopal@nokia.com MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 7bit >this mandates a multicast requirement in Forces? >As forces has a model where the CE-FE may be few hops away - it does >not make much value to run a muticast for such environment. Here also the destination is a few hops away only. multicast here refers to that done on an ethernet thereby saving on all the processing which needs to be done by the other processors not running that protocol. _________________________________________________________ Do You Yahoo!? Get your free @yahoo.com address at http://mail.yahoo.com 2001 Message-Id: Date: Sat, 8 Dec 2001 17:41:57 -0500 From: Stephen Suryaputra Subject: Re: (Reply) BGP in a distributed environment MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Manav, > Firstly, the ethernet interface that connects the multiple CEs is an "ethernet" in > the traditional sense of word. Secondly, and more importantly, the very behaviour > of RIP/OSPF suggests flooding routing information across its interfaces. And > this is what it will do in this case also when it has to convey information across > other CEs. This problem aggravates in case of BGP which is TCP unicast model > based. Hence we need to explore on this. If it is ethernet in traditional sense, then a legitimate IP interface must be configured before routing updates can be sent over the interface. Isn't that make the route selection process convoluted? I thought the ethernet interface we are talking about is a special interface inside of a router. Pls correct if I'm wrong. cheers, 2001 Message-Id: Date: Sat, 8 Dec 2001 15:19:05 +0000 From: Manav Bhatia Subject: Re: BGP in a distributed environment Comments: cc: todd.a.anderson@intel.com MIME-version: 1.0 Content-type: text/html; charset=EUC-KR Content-transfer-encoding: quoted-printable = =3Chtml=3E=3Cnobr=3E=3Cimg src=3Dhttp=3A//maildb=2Esamsung=2Ecom=3A8001/o= penconfirm=3Fparm=3Dmanav=5F=5Fmanav20011208205319895 width=3D=221=22 hei= ght=3D=221=22=3E=3C/nobr=3E =3Cfont size=3D3 face=3D=22=B1=BC=B8=B2=C3=BC=22=3E =3Cpre=3ETodd=2C Why do we need to encapsulate RIP or any=5Frouting=5Fprotocol packets via= the ForCES protocol =3F We simply allow RIP to multicast all of its packe= ts out on a virtual interface connecting to the backbone which will reach ev= ery other CE and the ones in the multicast group will accept the packet and p= rocess = it like any other RIP/OSPF/BGP/etc packet=2E Manav Todd wrote=3A I believe the presumption has been that when RIP multicasts its routing updates that those updates are tunneled to the FE via the ForCES protocol= and the FE then does the multicast on all of its =5Fexternal ports=5F=2E = If I were implementing a CE=2C I would therefore ignore any routing updates th= at appeared on my interfaces that weren=27t encapsulated with ForCES=2E I t= hink this gets us pretty much back to the normal case where routing protocols run between NE=27s but not inside them=2E Todd =3C/font=3E=3C/pre=3E=3C/html=3E 2001 Message-Id: Date: Sat, 8 Dec 2001 15:14:19 +0000 From: Manav Bhatia Subject: (Reply) BGP in a distributed environment Comments: cc: stephen@eecs.harvard.edu MIME-version: 1.0 Content-type: text/html; charset=EUC-KR Content-transfer-encoding: quoted-printable = =3Chtml=3E=3Cnobr=3E=3Cimg src=3Dhttp=3A//maildb=2Esamsung=2Ecom=3A8001/o= penconfirm=3Fparm=3Dmanav=5F=5Fmanav20011208204833787 width=3D=221=22 hei= ght=3D=221=22=3E=3C/nobr=3E =3Cfont size=3D3 face=3D=22=B1=BC=B8=B2=C3=BC=22=3E =3Cpre=3EStephen=2C As i understand and gather from what mareline says=2C this problem of con= veying control information across CEs does not arise with routing protocols like= RIP and OSPF=2E Why =3F Firstly=2C the ethernet interface that connects the multiple CEs is an =22= ethernet=22 in = the traditional sense of word=2E Secondly=2C and more importantly=2C the = very behaviour of RIP/OSPF suggests flooding routing information across its interfaces=2E= And this is what it will do in this case also when it has to convey informati= on across = other CEs=2E This problem aggravates in case of BGP which is TCP unicast = model = based=2E Hence we need to explore on this=2E Manav Stephen wrote=3A I have been following the discussion on and off=2E Apology if there is something that I missed=2E I think the problem still appears on RIP/OSPF because the ethernet that connects multiple CE must not be an interfaces in traditional sense=2C i=2E= e=2E the routing protocol doesn=27t take the interface into account during rou= te calculation=2E =3E Yes=2C very much=2E What is a CE+FE =3F Very crudely put=2C its a NE = or even worse=2C =3E a router by itself=2E Which goes to say that we can have many routers= inside a =3E router=2C where each router can/may be connected to the ethernet=2E N= ow we have =3E routing protocols running on each of these miniscule routers=2C and f= or the =3E sake of consistency we have to have all of these routers synchronizd = with =3E each other=2E I think it would be better if the routing protocol packets get broadcaste= d into all CE=2E Therefore=2C all routing protocol processes have approximately the same state=2E This especially trivial if multiple CEs are connected b= y ethernet=2E Comments=3F Cheers=2C =3C/font=3E=3C/pre=3E=3C/html=3E 2001 Message-Id: Date: Sat, 8 Dec 2001 12:46:23 -0500 From: Jamal Hadi Salim Organization: Znyx Networks Subject: Re: (Reply) BGP in a distributed environment Comments: To: Stephen Suryaputra MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: base64 DQpPbiBTYXR1cmRheSAwOCBEZWNlbWJlciAyMDAxIDE3OjQxLCBTdGVwaGVuIFN1cnlhcHV0cmEg d3JvdGU6DQo+SSB0aG91Z2h0IHRoZQ0KPiBldGhlcm5ldCBpbnRlcmZhY2Ugd2UgYXJlIHRhbGtp bmcgYWJvdXQgaXMgYSBzcGVjaWFsIGludGVyZmFjZSBpbnNpZGUgb2YgYQ0KPiByb3V0ZXIuIFBs cyBjb3JyZWN0IGlmIEknbSB3cm9uZy4NCj4NCg0KWW91IGFyZSByaWdodDsgaXQgc2VlbXMgYSBs b3Qgb3RoZXIgcG9zdGluZ3Mgd2VyZSBtaXNzaW5nIHRoaXMgc3VidGxlIHBvaW50IA0KKGhlbmNl IG1ha2luZyByZWZlcmVuY2VzIHRvIFRDUCwgQkdQIGV0YykNCg0KY2hlZXJzLA0KamFtYWwNCg== 2001 Message-Id: Date: Sat, 8 Dec 2001 11:21:16 -0500 From: "Martin, Olivier" Subject: RE : Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Manav, It should be encapsulated because the ethernet network between the FE and CE is not a traditionnal network in the sense that the NE has to be seen as an entire router, and should there be any other router on this FE-CE ethernet segment, it can't see the updates that the FE got. Typically, this ethernet segment would be inside a proprietary box, but we have to think of it as an open standard, and as such, be able to separate the FE from the CE using any physical media. Regards, Olivier -----Message d'origine----- De : Manav Bhatia [mailto:manav@samsung.com]=20 Envoy=E9 : 8 d=E9cembre, 2001 10:19 =C0 : FORCES@PEACH.EASE.LSOFT.COM Objet : Re: BGP in a distributed environment =20 Todd, Why do we need to encapsulate RIP or any_routing_protocol packets via the ForCES protocol ? We simply allow RIP to multicast all of its packets out on a virtual interface connecting to the backbone which will reach every other CE and the ones in the multicast group will accept the packet and process=20 it like any other RIP/OSPF/BGP/etc packet. Manav Todd wrote: I believe the presumption has been that when RIP multicasts its routing updates that those updates are tunneled to the FE via the ForCES protocol and the FE then does the multicast on all of its _external ports_. If I were implementing a CE, I would therefore ignore any routing updates that appeared on my interfaces that weren't encapsulated with ForCES. I think this gets us pretty much back to the normal case where routing protocols run between NE's but not inside them. Todd 2001 Message-Id: Date: Sat, 8 Dec 2001 10:23:33 -0500 From: "Gopal Ram (NRC/Boston)" Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: multipart/alternative; boundary="----_=_NextPart_001_01C17FFC.4C5D6174" This is a multi-part message in MIME format. ------_=_NextPart_001_01C17FFC.4C5D6174 Content-Type: text/plain; charset="ks_c_5601-1987" Content-Transfer-Encoding: quoted-printable [] Hello Manav , See my inline comments. -----Original Message----- From: ext Manav Bhatia [mailto:manav@samsung.com] Sent: Saturday, December 08, 2001 10:19 AM To: FORCES@PEACH.EASE.LSOFT.COM Subject: Re: BGP in a distributed environment =20 =20 Todd, Why do we need to encapsulate RIP or any_routing_protocol packets via the ForCES protocol ? We simply allow RIP to multicast all of its packets out on a virtual interface connecting to the backbone which will reach every other CE and the ones in the multicast group will accept the packet and process=20 it like any other RIP/OSPF/BGP/etc packet. []=20 this mandates a multicast requirement in Forces? As forces has a model where the CE-FE may be few hops away - it does not make much value to run a muticast for such environment. Regards Ramg =20 Manav Todd wrote: I believe the presumption has been that when RIP multicasts its routing updates that those updates are tunneled to the FE via the ForCES protocol and the FE then does the multicast on all of its _external ports_. If I were implementing a CE, I would therefore ignore any routing updates that appeared on my interfaces that weren't encapsulated with ForCES. I think this gets us pretty much back to the normal case where routing protocols run between NE's but not inside them. Todd ------_=_NextPart_001_01C17FFC.4C5D6174 Content-Type: text/html; charset="ks_c_5601-1987" Content-Transfer-Encoding: quoted-printable

[<Ramg>]  Hello Manav=20 ,
See my inline=20 comments.
 -----Original=20 Message-----
From: ext Manav Bhatia=20 [mailto:manav@samsung.com]
Sent: Saturday, December 08, 2001 = 10:19=20 AM
To: FORCES@PEACH.EASE.LSOFT.COM
Subject: Re: = BGP in a=20 distributed environment

Todd,
Why do we need to encapsulate RIP or any_routing_protocol packets via
the ForCES protocol ? We simply allow RIP to multicast all of its =
packets
out on a virtual interface connecting to the backbone which will reach =
every
other CE and the ones in the multicast group will accept the packet and =
process=20
it like any other RIP/OSPF/BGP/etc packet.

[<Ramg>] 
this mandates a multicast requirement in = Forces?
As forces has a model where = the CE-FE may be few hops away - it does not make much value  to = run a muticast for such environment.
Regards
Ramg
 
Manav Todd wrote: I believe the presumption has been that when RIP multicasts its routing updates that those updates are tunneled to the FE via the ForCES = protocol and the FE then does the multicast on all of its _external ports_. If I were implementing a CE, I would therefore ignore any routing updates = that appeared on my interfaces that weren't encapsulated with ForCES. I = think this gets us pretty much back to the normal case where routing protocols run between NE's but not inside them. Todd
------_=_NextPart_001_01C17FFC.4C5D6174-- 2001 Message-Id: Date: Fri, 7 Dec 2001 16:25:32 EST From: "\"" Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="US-ASCII" Content-Transfer-Encoding: 7bit David, I agree. I have exchanged some idea's with Mark on this subject as well as with Todd. It is a good starting point to address these issues. Bob Hines Robert Hines Inc. Burbank, Ca. 2001 Message-Id: Date: Fri, 7 Dec 2001 13:59:48 -0500 From: "Martin, Olivier" Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Is the CE-FE relationship supposed to be ultimately able to be manufacturer-independant ? When in the requirements, it is question of having the close-proximity between the CE and FE, even if it says that it could be one hop away, could the CE be 25 hops, the whole Internet, appart from the FE ? If so, there definitely is a need for the ForCES protocol on the CE-CE / CE-FE link to be defined. I agree with you big time, Todd, on the fact that conflicts have to be resolved by whatever CE(s) are controlling the FEs. This an absolute must as you don't want two FEs to take 2 different decisions, because of timing events or whatever could happen in between (delay on hops...) that would make this happen. Should we think about defining a ForCES encapsulation ? And transport mecanism ? Even if it should not be defined by the ForCES WG, we could take the time to create a WG for this purpose ? Sorry if I write this as being a simple thing (to form another WG), as I'm new to the community and I don't know what the procedures are to create a WG. Olivier -----Message d'origine----- De : Anderson, Todd A [mailto:todd.a.anderson@intel.com] Envoy=E9 : 7 d=E9cembre, 2001 13:46 =C0 : FORCES@PEACH.EASE.LSOFT.COM Objet : Re: BGP in a distributed environment > -----Original Message----- > From: Tom Petch [mailto:nwnetworks@dial.pipex.com] > Sent: Saturday, December 07, 2002 1:55 AM > To: FORCES@PEACH.EASE.LSOFT.COM > Subject: Re: [FORCES] BGP in a distributed environment > But what should also be in scope is the behaviour of an FE > when it receives a control object from a CE that > substantially changes a control object it received from > another CE, so if an OSPF CE forwarding entry is replaced by > a BGP or NAT CE entry, then .. well the behaviour is > defined. We may not define the behaviour of FE to the level > of Cisco's Admin Distance but we should at least recognise > that this happens and identify behaviour(s) eg new replaces > old and old's CE is notified so the CE can get together and > resolve the issue. If CEs have to get together to resolve potential conflicts anyway, why not have them do so before they modify the forwarding behavior of the FE? If CEs resolve conflicts before hand, it seems to me that we can eliminate the need for any registration/callback mechanism from the FE to CE. I'm a big fan of keeping FEs as simple as possible and the policy that you do whatever any of your CEs tell you to do based on the order in which those instructions arrive seems to me the simplest possible arrangement. Todd 2001 Message-Id: Date: Fri, 7 Dec 2001 12:59:35 -0800 From: "Putzolu, David" Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" > Is the CE-FE relationship supposed to be ultimately able to be > manufacturer-independant ? When in the requirements, it is > question of > having the close-proximity between the CE and FE, even if it says that > it could be one hop away, could the CE be 25 hops, the whole Internet, > appart from the FE ? If so, there definitely is a need for the ForCES > protocol on the CE-CE / CE-FE link to be defined. An individual contribution has been submitted for ForCES applicability that touches on this - worth a read, it is on agenda for Monday: http://www.ietf.org/internet-drafts/draft-crouch-forces-applicability-00.txt Cheers, David 2001 Message-Id: Date: Fri, 7 Dec 2001 12:43:13 -0500 From: "Martin, Olivier" Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Sorry I did not read all papers yet other than the 2 currently visible drafts on the ietf's WG page, but is there any plans about specifying would be the ForCES protocol ? Defining a way of tunelling data ? The medium between the CEs and FEs has been specified as being anything, from a proprietary bus to a ethernet/ATM bus. There needs to be a frame format, a little like the one in the Linux Netlink. But again, this may all have been done yet in another paper and if it's the case, I apologise for not having taken the time to read it. Olivier -----Message d'origine----- De : Anderson, Todd A [mailto:todd.a.anderson@intel.com] Envoy=E9 : 7 d=E9cembre, 2001 12:34 =C0 : FORCES@PEACH.EASE.LSOFT.COM Objet : Re: BGP in a distributed environment I believe the presumption has been that when RIP multicasts its routing updates that those updates are tunneled to the FE via the ForCES protocol and the FE then does the multicast on all of its _external ports_. If I were implementing a CE, I would therefore ignore any routing updates that appeared on my interfaces that weren't encapsulated with ForCES. I think this gets us pretty much back to the normal case where routing protocols run between NE's but not inside them. Todd -----Original Message----- From: Mareline Sheldon [mailto:marelines@yahoo.com] Sent: Thursday, December 06, 2001 10:36 PM To: FORCES@PEACH.EASE.LSOFT.COM Subject: Re: [FORCES] BGP in a distributed environment Actually the problem does not arise with the other routing protocols like RIP/OSPF because the inherent behaviour of such protocols is that they flood out information indiscriminatly out on all the interfaces configured for the protocol. So if we have multiple CEs connected on an ethernet then that flood will send the routing information to the other CEs which can accept the packet and process it like wise. This is not possible with BGP which cannot be broadcast since it runs over TCP. >Is it possible to have multiple CEs in a single router? Yes, very much. What is a CE+FE ? Very crudely put, its a NE or even worse, a router by itself. Which goes to say that we can have many routers inside a router, where each router can/may be connected to the ethernet. Now we have routing protocols running on each of these miniscule routers, and for the sake of consistency we have to have all of these routers synchronizd with each other. Hence the whole problem. Thanks and Regards, Mareline Margaret wrote : Date: Thu, 6 Dec 2001 14:55:40 -0500 From: Margaret Wasserman Subject: Re: BGP in a distributed environment Hi Jamal, If I understood correctly, Mareline was discussing the best way to distribute BGP across multiple CEs. I suppose that the same problem might also arise when trying to distribute OSPF and/or RIP across multiple CEs? Is it possible to have multiple CEs in a single router? If so, should this working group define a protocol for how control information (such as routing information) is distributed between those CEs? Or should we just focus on ME <-> CE <-> FE interactions? Margaret At 12:28 PM 12/6/01 , Jamal Hadi Salim wrote: >I am missing something here; why IDR? Why should BGP be a specail case? >This is why we have ForCES to begin with. You could have a single BGP compute >node which could serve multiple NEs as s/he suggest in his 2) >Removing that single point of failure is out of the scope of ForCES. If >within a reliable environment (such as likely within the scope a single box), >multicast for distributing the BGP compute results to the NE is a fair use of >the still to be defined ForCES protocol. TCP/SCTP only comes into play when >going across multiple hops. This, i thought was clear in the requirements doc. > >just my $.02 cents .ca > _________________________________________________________ Do You Yahoo!? Get your free @yahoo.com address at http://mail.yahoo.com 2001 Message-Id: Date: Fri, 7 Dec 2001 12:25:04 +0530 From: Mareline Sheldon Subject: Re: BGP in a distributed environment Comments: To: todd.a.anderson@intel.com MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 7bit Todd, Have one central arbiter to decide which all routes to finally select is not a good option. In fact this was one idea in my original mail. Drawbacks: (1) You have a single point of failure. For resilience we can put up a backup processor which will take over whenever the primary fails but this requires additional complexity in keeping them synchronized, etc. Additional overhead will be dynamically choosing the "designated' processor [borrowed the term from OSPF!]. Moreover injecting the full BGP feed into one processor is not a very good idea, especially when you may have multiple CEs dealing with complete BGP tables (core) In my view, each CE should have its routes which traditionally are given to FEs. With multiple CEs and FEs around we must have a mechansim wherein the CEs are able to announce their routes to each other. The problem does not end here. It infact just starts .. Each CE not only needs to gives its forwarding routes but it needs to send its all the route entries togethar with all its attributes. To cite an example, a CE will have to advertise the entire BGP route with all its path attributes to the other CE so that it may then i turn re-run its decision process and come up with new routes if so required. We must thus evolve standards for inter CE communication also ! Your views please ? Regards, Mareline S. You wrote: Date: Thu, 6 Dec 2001 11:34:32 -0800 From: "Anderson, Todd A" Subject: Re: BGP in a distributed environment Content-Type: text/plain; charset="iso-8859-1" People have brought up this idea before of having different routing protocols running on different CEs but the problem comes when those protocols want to apply what they have learned. Essentially, you need some arbiter that decides which protocol's routes take precedence. I don't think you want FEs to have the sort of knowledge necessary to make this decision. My view is that you need a centralized route arbiter in one of the CEs to which all protocols send their results and the arbiter produces a unified table that it then sends diffs to the FEs. However, how the CEs communicate with each other or the arbiter I think is currently out of scope of ForCES and the requirements document places no restrictions on CE to CE communication. Todd _________________________________________________________ Do You Yahoo!? Get your free @yahoo.com address at http://mail.yahoo.com 2001 Message-Id: Date: Fri, 7 Dec 2001 12:05:43 +0530 From: Mareline Sheldon Subject: Re: BGP in a distributed environment Comments: To: mrw@windriver.com MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 7bit Actually the problem does not arise with the other routing protocols like RIP/OSPF because the inherent behaviour of such protocols is that they flood out information indiscriminatly out on all the interfaces configured for the protocol. So if we have multiple CEs connected on an ethernet then that flood will send the routing information to the other CEs which can accept the packet and process it like wise. This is not possible with BGP which cannot be broadcast since it runs over TCP. >Is it possible to have multiple CEs in a single router? Yes, very much. What is a CE+FE ? Very crudely put, its a NE or even worse, a router by itself. Which goes to say that we can have many routers inside a router, where each router can/may be connected to the ethernet. Now we have routing protocols running on each of these miniscule routers, and for the sake of consistency we have to have all of these routers synchronizd with each other. Hence the whole problem. Thanks and Regards, Mareline Margaret wrote : Date: Thu, 6 Dec 2001 14:55:40 -0500 From: Margaret Wasserman Subject: Re: BGP in a distributed environment Hi Jamal, If I understood correctly, Mareline was discussing the best way to distribute BGP across multiple CEs. I suppose that the same problem might also arise when trying to distribute OSPF and/or RIP across multiple CEs? Is it possible to have multiple CEs in a single router? If so, should this working group define a protocol for how control information (such as routing information) is distributed between those CEs? Or should we just focus on ME <-> CE <-> FE interactions? Margaret At 12:28 PM 12/6/01 , Jamal Hadi Salim wrote: >I am missing something here; why IDR? Why should BGP be a specail case? >This is why we have ForCES to begin with. You could have a single BGP compute >node which could serve multiple NEs as s/he suggest in his 2) >Removing that single point of failure is out of the scope of ForCES. If >within a reliable environment (such as likely within the scope a single box), >multicast for distributing the BGP compute results to the NE is a fair use of >the still to be defined ForCES protocol. TCP/SCTP only comes into play when >going across multiple hops. This, i thought was clear in the requirements doc. > >just my $.02 cents .ca > _________________________________________________________ Do You Yahoo!? Get your free @yahoo.com address at http://mail.yahoo.com 2001 Message-Id: Date: Fri, 7 Dec 2001 10:46:16 -0800 From: "Anderson, Todd A" Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" > -----Original Message----- > From: Tom Petch [mailto:nwnetworks@dial.pipex.com] > Sent: Saturday, December 07, 2002 1:55 AM > To: FORCES@PEACH.EASE.LSOFT.COM > Subject: Re: [FORCES] BGP in a distributed environment > But what should also be in scope is the behaviour of an FE > when it receives a control object from a CE that > substantially changes a control object it received from > another CE, so if an OSPF CE forwarding entry is replaced by > a BGP or NAT CE entry, then .. well the behaviour is > defined. We may not define the behaviour of FE to the level > of Cisco's Admin Distance but we should at least recognise > that this happens and identify behaviour(s) eg new replaces > old and old's CE is notified so the CE can get together and > resolve the issue. If CEs have to get together to resolve potential conflicts anyway, why not have them do so before they modify the forwarding behavior of the FE? If CEs resolve conflicts before hand, it seems to me that we can eliminate the need for any registration/callback mechanism from the FE to CE. I'm a big fan of keeping FEs as simple as possible and the policy that you do whatever any of your CEs tell you to do based on the order in which those instructions arrive seems to me the simplest possible arrangement. Todd 2002 Message-Id: Date: Sat, 7 Dec 2002 09:54:41 -0000 From: Tom Petch Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 7bit Agree that CE-CE is out of scope of ForCES (but needs doing somewhere!) Agree that FE should have a record of why it is doing what it is, eg where it received a given IP forwarding entry from. But what should also be in scope is the behaviour of an FE when it receives a control object from a CE that substantially changes a control object it received from another CE, so if an OSPF CE forwarding entry is replaced by a BGP or NAT CE entry, then .. well the behaviour is defined. We may not define the behaviour of FE to the level of Cisco's Admin Distance but we should at least recognise that this happens and identify behaviour(s) eg new replaces old and old's CE is notified so the CE can get together and resolve the issue. If we don't do this then we severely limit the scope for multiple CE eg to one CE for each element such as forwarding engine. I am also conscious that the scope of ForCES is to allow elements to be multiple hops apart so anything that depends on multi- or broad- cast is too restrictive. (And I think that is worth more than one cent :-) Tom Petch, Network Consultant nwnetworks@dial.pipex.com 2001 Message-Id: Date: Fri, 7 Dec 2001 09:34:07 -0800 From: "Anderson, Todd A" Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" I believe the presumption has been that when RIP multicasts its routing updates that those updates are tunneled to the FE via the ForCES protocol and the FE then does the multicast on all of its _external ports_. If I were implementing a CE, I would therefore ignore any routing updates that appeared on my interfaces that weren't encapsulated with ForCES. I think this gets us pretty much back to the normal case where routing protocols run between NE's but not inside them. Todd -----Original Message----- From: Mareline Sheldon [mailto:marelines@yahoo.com] Sent: Thursday, December 06, 2001 10:36 PM To: FORCES@PEACH.EASE.LSOFT.COM Subject: Re: [FORCES] BGP in a distributed environment Actually the problem does not arise with the other routing protocols like RIP/OSPF because the inherent behaviour of such protocols is that they flood out information indiscriminatly out on all the interfaces configured for the protocol. So if we have multiple CEs connected on an ethernet then that flood will send the routing information to the other CEs which can accept the packet and process it like wise. This is not possible with BGP which cannot be broadcast since it runs over TCP. >Is it possible to have multiple CEs in a single router? Yes, very much. What is a CE+FE ? Very crudely put, its a NE or even worse, a router by itself. Which goes to say that we can have many routers inside a router, where each router can/may be connected to the ethernet. Now we have routing protocols running on each of these miniscule routers, and for the sake of consistency we have to have all of these routers synchronizd with each other. Hence the whole problem. Thanks and Regards, Mareline Margaret wrote : Date: Thu, 6 Dec 2001 14:55:40 -0500 From: Margaret Wasserman Subject: Re: BGP in a distributed environment Hi Jamal, If I understood correctly, Mareline was discussing the best way to distribute BGP across multiple CEs. I suppose that the same problem might also arise when trying to distribute OSPF and/or RIP across multiple CEs? Is it possible to have multiple CEs in a single router? If so, should this working group define a protocol for how control information (such as routing information) is distributed between those CEs? Or should we just focus on ME <-> CE <-> FE interactions? Margaret At 12:28 PM 12/6/01 , Jamal Hadi Salim wrote: >I am missing something here; why IDR? Why should BGP be a specail case? >This is why we have ForCES to begin with. You could have a single BGP compute >node which could serve multiple NEs as s/he suggest in his 2) >Removing that single point of failure is out of the scope of ForCES. If >within a reliable environment (such as likely within the scope a single box), >multicast for distributing the BGP compute results to the NE is a fair use of >the still to be defined ForCES protocol. TCP/SCTP only comes into play when >going across multiple hops. This, i thought was clear in the requirements doc. > >just my $.02 cents .ca > _________________________________________________________ Do You Yahoo!? Get your free @yahoo.com address at http://mail.yahoo.com 2001 Message-Id: Date: Fri, 7 Dec 2001 08:37:06 -0500 From: Stephen Suryaputra Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII I have been following the discussion on and off. Apology if there is something that I missed. On Fri, 7 Dec 2001, Mareline Sheldon wrote: > Actually the problem does not arise with the other routing protocols like > RIP/OSPF because the inherent behaviour of such protocols is that they flood > out information indiscriminatly out on all the interfaces configured for the > protocol. So if we have multiple CEs connected on an ethernet then that > flood will send the routing information to the other CEs which can accept > the packet and process it like wise. > > This is not possible with BGP which cannot be broadcast since it runs over > TCP. I think the problem still appears on RIP/OSPF because the ethernet that connects multiple CE must not be an interfaces in traditional sense, i.e. the routing protocol doesn't take the interface into account during route calculation. > Yes, very much. What is a CE+FE ? Very crudely put, its a NE or even worse, > a router by itself. Which goes to say that we can have many routers inside a > router, where each router can/may be connected to the ethernet. Now we have > routing protocols running on each of these miniscule routers, and for the > sake of consistency we have to have all of these routers synchronizd with > each other. I think it would be better if the routing protocol packets get broadcasted into all CE. Therefore, all routing protocol processes have approximately the same state. This especially trivial if multiple CEs are connected by ethernet. Comments? Cheers, > > Hence the whole problem. > > Thanks and Regards, > Mareline > > Margaret wrote : > > Date: Thu, 6 Dec 2001 14:55:40 -0500 > From: Margaret Wasserman > Subject: Re: BGP in a distributed environment > Hi Jamal, > > > If I understood correctly, Mareline was discussing the best way to > distribute BGP across multiple CEs. I suppose that the same problem > might also arise when trying to distribute OSPF and/or RIP across > multiple CEs? > > > Is it possible to have multiple CEs in a single router? If so, should > this working group define a protocol for how control information (such > as routing information) is distributed between those CEs? Or should > we just focus on ME <-> CE <-> FE interactions? > > Margaret > > At 12:28 PM 12/6/01 , Jamal Hadi Salim wrote: > > >I am missing something here; why IDR? Why should BGP be a specail case? > >This is why we have ForCES to begin with. You could have a single BGP > compute > >node which could serve multiple NEs as s/he suggest in his 2) > >Removing that single point of failure is out of the scope of ForCES. If > >within a reliable environment (such as likely within the scope a single > box), > >multicast for distributing the BGP compute results to the NE is a fair use > of > >the still to be defined ForCES protocol. TCP/SCTP only comes into play when > >going across multiple hops. This, i thought was clear in the requirements > doc. > > > >just my $.02 cents .ca > > > > > _________________________________________________________ > Do You Yahoo!? > Get your free @yahoo.com address at http://mail.yahoo.com > > 2001 Message-Id: Date: Fri, 7 Dec 2001 07:45:25 -0500 From: "Martin, Olivier" Subject: RE : Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Mareline, In the discussions we had in the past 24h, I wonder if Todd (and other) are right when they say that this ForCES WG is only there to say which function should be performed by which part of the NE. Should how they do it (e.-g., CE-CE communications) be a matter of some other document ? Should this CE-CE communitation be part of the ForCES requirement draft ? I think that it should be part of it. There could be a frame format, defined, which would describe how data would be exchanged on the ForCES network, or proprietary bus. Since a NE is supposed to be view as a single entity from the exterior world, I don't think that the router could allow itself to have different calculations forwarded out to different FEs, whatever the results of the calculations are (routes for example). The behavior of a FE offering the same service as another FE should be the same, whichever CE controls it. I think the specific case of BGP should be discussed on another document too. But I'm not sure how this could be done specifically as this depends on the acceptation of the ForCES drafts, and its evolution to a RFC and it might still have some time to wait before this happens. Any suggestions ? Olivier -----Message d'origine----- De : Mareline Sheldon [mailto:marelines@yahoo.com]=20 Envoy=E9 : 7 d=E9cembre, 2001 01:55 =C0 : FORCES@PEACH.EASE.LSOFT.COM Objet : Re: BGP in a distributed environment Todd, Have one central arbiter to decide which all routes to finally select is not a good option. In fact this was one idea in my original mail. Drawbacks: (1) You have a single point of failure. For resilience we can put up a backup processor which will take over whenever the primary fails but this requires additional complexity in keeping them synchronized, etc. Additional overhead will be dynamically choosing the "designated' processor [borrowed the term from OSPF!]. Moreover injecting the full BGP feed into one processor is not a very good idea, especially when you may have multiple CEs dealing with complete BGP tables (core) In my view, each CE should have its routes which traditionally are given to FEs. With multiple CEs and FEs around we must have a mechansim wherein the CEs are able to announce their routes to each other. The problem does not end here. It infact just starts .. Each CE not only needs to gives its forwarding routes but it needs to send its all the route entries togethar with all its attributes. To cite an example, a CE will have to advertise the entire BGP route with all its path attributes to the other CE so that it may then i turn re-run its decision process and come up with new routes if so required. We must thus evolve standards for inter CE communication also ! Your views please ? Regards, Mareline S. You wrote: Date: Thu, 6 Dec 2001 11:34:32 -0800 From: "Anderson, Todd A" Subject: Re: BGP in a distributed environment Content-Type: text/plain; charset=3D"iso-8859-1" People have brought up this idea before of having different routing protocols running on different CEs but the problem comes when those protocols want to apply what they have learned. Essentially, you need some arbiter that decides which protocol's routes take precedence. I don't think you want FEs to have the sort of knowledge necessary to make this decision. My view is that you need a centralized route arbiter in one of the CEs to which all protocols send their results and the arbiter produces a unified table that it then sends diffs to the FEs. However, how the CEs communicate with each other or the arbiter I think is currently out of scope of ForCES and the requirements document places no restrictions on CE to CE communication. Todd _________________________________________________________ Do You Yahoo!? Get your free @yahoo.com address at http://mail.yahoo.com 2001 Message-Id: Date: Thu, 6 Dec 2001 22:25:32 -0500 From: Jamal Hadi Salim Organization: Znyx Networks Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: base64 DQpNYXJnYXJldCwNCg0KT24gVGh1cnNkYXkgMDYgRGVjZW1iZXIgMjAwMSAxNDo1NSwgTWFyZ2Fy ZXQgV2Fzc2VybWFuIHdyb3RlOg0KPiBIaSBKYW1hbCwNCj4NCj4gSWYgSSB1bmRlcnN0b29kIGNv cnJlY3RseSwgTWFyZWxpbmUgd2FzIGRpc2N1c3NpbmcgdGhlIGJlc3Qgd2F5IHRvDQo+IGRpc3Ry aWJ1dGUgQkdQIGFjcm9zcyBtdWx0aXBsZSBDRXMuICBJIHN1cHBvc2UgdGhhdCB0aGUgc2FtZSBw cm9ibGVtDQo+IG1pZ2h0IGFsc28gYXJpc2Ugd2hlbiB0cnlpbmcgdG8gZGlzdHJpYnV0ZSBPU1BG IGFuZC9vciBSSVAgYWNyb3NzDQo+IG11bHRpcGxlIENFcz8NCg0Kbm9kOyANCg0KPg0KPiBJcyBp dCBwb3NzaWJsZSB0byBoYXZlIG11bHRpcGxlIENFcyBpbiBhIHNpbmdsZSByb3V0ZXI/ICBJZiBz bywgc2hvdWxkDQo+IHRoaXMgd29ya2luZyBncm91cCBkZWZpbmUgYSBwcm90b2NvbCBmb3IgaG93 IGNvbnRyb2wgaW5mb3JtYXRpb24gKHN1Y2gNCj4gYXMgcm91dGluZyBpbmZvcm1hdGlvbikgaXMg ZGlzdHJpYnV0ZWQgYmV0d2VlbiB0aG9zZSBDRXM/ICBPciBzaG91bGQNCj4gd2UganVzdCBmb2N1 cyBvbiBNRSA8LT4gQ0UgPC0+IEZFIGludGVyYWN0aW9ucz8NCg0KQ0U8LT5DRSBpcyBkZWZpbmV0 ZWx5IG91dCBzY29wZS4gQnV0IGkgdGhpbmsgdGhhdCB0aGVyZSBzaG91bGQgYmUgYSB3YXkgdG8g DQppZGVudGlmeSB3aG8gKENFKSBpbnN0YWxsZWQgd2hhdCBvbiB0aGUgRkUuIEkgdGhpbmsgdGhh dCBtZWNoYW5pc20gYXQgbGVhc3QgDQpuZWVkcyB0byBiZSBwYXJ0IG9mIHRoZSBGb3JDRVMgcHJv dG9jb2wgLi4uDQoNCmNoZWVycywNCmphbWFsDQo= 2001 Message-Id: Date: Thu, 6 Dec 2001 15:52:28 EST From: "\"" Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="US-ASCII" Content-Transfer-Encoding: 7bit Thanks Mareline, This has to be the most posting I have seen on this list ever. But I do feel as well that this is an implementation issue and would have to agree with Todd that this subject nonetheless out of the ForCes scope. I refer to my previous query regarding the draft-crouch-forces-applicability-00.txt. >>Alan >>When a case arises that a multiple CE configured box sharing outlined defined multiple FE information on a common bus, Do you feel that a situation may arise where a "Master CE" definition will be needed? Or will all CE's be treated as equal? Trying to maintain this many possible (one example QoS flows) on a Terabit Ether switch without a master slave relationships may prove very challenging. This out of the ForCES scope and more dependent on say IFIX proposal to act as arbitration between multiple CE's as in (FE<->CE< ->ME).<< my 1 cent we might be in a recession Bob Robert Hines Inc. Burbank Ca. 2001 Message-Id: Date: Thu, 6 Dec 2001 15:48:18 +0530 From: Mareline Sheldon Subject: BGP in a distributed environment MIME-Version: 1.0 Content-Type: multipart/mixed; boundary="----=_NextPart_000_04CC_01C17E6D.6CB91B10" This is a multi-part message in MIME format. ------=_NextPart_000_04CC_01C17E6D.6CB91B10 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 7bit Hi All, The mail is attached for it has a diagram which would be distorted otherwise. regards, mareline s ------=_NextPart_000_04CC_01C17E6D.6CB91B10 Content-Type: application/msword; name="bgp_distribution.doc" Content-Transfer-Encoding: quoted-printable Content-Disposition: attachment; filename="bgp_distribution.doc" Hi, A network element is composed of numerous logically separated entities = that cooperate to provide a given functionality (such as a routing or IP = switching) and yet appear as a normal integrated network element to = external entities. Two primary types of network element components = exist: control-plane components and forwarding-plane components (ASIC). = A standard set of mechanisms for connecting these components would = provide increased scalability and allow the control and forwarding = planes to evolve independently thus promoting faster innovation. Going a step further, we conglomerate these various network elements and = put them on an Ethernet to form a truly integrated network element [read = as, terabit router] composed of multiple packet processing engines = interconnected to form a logical whole. [Ref to the figure below] -------------------------------------------------------------------------= ---- | -------------------------------- = ------------------------------- | | | NE1 = | | NE2 | | | | = ------------------------------------------- | | | | = | | | | | | = | | ------------- | | ------------- = | | | | | CE | | | | CE = | | | | | ------------- | | = ------------- | | | | / \ | = | / \ | | | | / \ = | | / \ | | | | / = \ | | / \ | | | | / = \ | | --------- --------- | | | | = ----------- ----------- | | | FE | | FE | | | = | | | FE | | FE | | | ---------- --------- = | | | | ----------- ----------- | | | | | = | | | | | | | | | | | | | | | | | | | = | | | | | | | | | | | | | | | | | = | | | | | | | | | | | | | | | | | | | = | ------------------------------- | | = -------------------------------- | | | | | | = | | | | | | | | | | | | | | = | | | = |------------------------------------------------------------------------= ---- | | | | | | | | | | | | | | ^ ^ | | BGP BGP NE Network element CE Control element=20 FE Forwarding element Traditional routers implement routing protocols on a centralized = processor, which has visibility to all the interfaces associated with = the router. Such a system differs from a traditional router in a sense = that both the control and data processing are distributed across the = different processors. Suppose we are running BGP on one of the interfaces of one network = element [a.k.a. router]. Whatever routes it learns must be advertised to = all the remaining network elements. This can be done via some = redistribution mechanism within the router. The problem comes in when we = are running BGP on another interface belonging to some other NE (Say 2 = in our case). Now whatever routes NE1 learns, must be passed along to = NE2 with *ALL* its attributes so that it can run its decision process = and come up with the optimal routes. The same holds true for NE1. It = must have all the routes along with the complete attributes from NE2 so = that it can run its decision process. Once they have a consistent table, = the routes can be leaked into the remaining network elements [routers]. = This problem aggravates when we have network elements [routers] of the = order of O (1000) What are the options left with us to ensure that the BGP tables across = the NEs remain synchronized. (1) We cant have each router peering up with the other router to form a = full IBGP mesh as this would lead to scalability issues. (2) We can have a central server to which all BGP learnt routes are = sent. There it runs its decision process and there propagates those = selected routes to every other router. This creates a single point of = failure. Moreover, with the full BGP routing table being injected there = is going to be a considerable load on the designated router/processor. (3) We can have hierarchical route reflectors within the router but it = is also not scalable. Secondly, there would be a lot of complexities = involved in the selection of server and the clients, etc. (4) Hack the BGP protocol and making it TCP independent, broadcasting = all the BGP packets within the consolidated router for within this = integrated networking element all routers running BGP are peers. = Reliability is at stake (which we don't ever bother with TCP!).=20 Is there any thought in the working group on standardizing such = distributed BGP route computations? Regards, Manav Bhatia ------=_NextPart_000_04CC_01C17E6D.6CB91B10-- _________________________________________________________ Do You Yahoo!? Get your free @yahoo.com address at http://mail.yahoo.com 2001 Message-Id: Date: Thu, 6 Dec 2001 14:55:40 -0500 From: Margaret Wasserman Subject: Re: BGP in a distributed environment Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Hi Jamal, If I understood correctly, Mareline was discussing the best way to distribute BGP across multiple CEs. I suppose that the same problem might also arise when trying to distribute OSPF and/or RIP across multiple CEs? Is it possible to have multiple CEs in a single router? If so, should this working group define a protocol for how control information (such as routing information) is distributed between those CEs? Or should we just focus on ME <-> CE <-> FE interactions? Margaret At 12:28 PM 12/6/01 , Jamal Hadi Salim wrote: >I am missing something here; why IDR? Why should BGP be a specail case? >This is why we have ForCES to begin with. You could have a single BGP compute >node which could serve multiple NEs as s/he suggest in his 2) >Removing that single point of failure is out of the scope of ForCES. If >within a reliable environment (such as likely within the scope a single box), >multicast for distributing the BGP compute results to the NE is a fair use of >the still to be defined ForCES protocol. TCP/SCTP only comes into play when >going across multiple hops. This, i thought was clear in the requirements doc. > >just my $.02 cents .ca > >cheers, >jamal 2001 Message-Id: Date: Thu, 6 Dec 2001 13:32:13 -0500 From: "Martin, Olivier" Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Hi everyone, I might be mistaking, but the same issue would arise with any IGP too. There could be a different central CE per routing protocol instance, that would send the forwarding tables along with the results of its calculation to the FEs, which in turn would use that information to populate their own forwarding tables in order to properly send the egress traffic to the proper interface / FE. The same information should be sent out to other CEs in the system, so if there are dependencies between some IP service and another one (such as the relationship between BGP and the IGP within the AS), the CEs would be aware of it. All of this communication should be using the mecanism described in the requirement draft, over a the [RFC2914]-compliant protocol. =20 This would create the possibility of having N CEs, each taking care of a particular function within the NE and since these entities are using GP CPUs, they could take over the process of some failed CE, using some VRRP or similar functionnality. =20 My 2 cents.ca too ;) !=20 Cheers, Olivier -----Message d'origine----- De : Margaret Wasserman [mailto:mrw@windriver.com] Envoy=E9 : 6 d=E9cembre, 2001 12:05 =C0 : FORCES@PEACH.EASE.LSOFT.COM Objet : Re: BGP in a distributed environment Hi Mareline, >(1) We cant have each router peering up with the other router to form a >full IBGP mesh as this would lead to scalability issues. Are there any scalability issues introduced by IBGP that are not inherent in distributing this information between a mesh of nodes using any unicast communications channel? If so, you should probably raise this issue with the IDR working group (responsible for BGP). >(2) We can have a central server to which all BGP learnt routes are sent. >There it runs its decision process and there propagates those selected routes >to every other router. This creates a single point of failure. Moreover, >with the full BGP routing table being injected there is going to be a >considerable load on the designated router/processor. How would this be different than running a single BGP on a single control element and having the distributed nodes forward all BGP traffic to that control element? In other words, if the BGP protocol is not well-suited to being distributed across multiple CEs, why distribute it? >(4) Hack the BGP protocol and making it TCP independent, broadcasting all >the BGP packets within the consolidated router for within this integrated >networking element all routers running BGP are peers. Reliability is at >stake (which we don't ever bother with TCP!). A multicast version of IBGP is an interesting idea. It would probably scale well when all IBGP nodes within a single group are attached to a single ethernet (as in your example). How would it scale if the IBGP nodes were distributed throughout a complex routing domain? >Is there any thought in the working group on standardizing such >distributed BGP route computations? In my opinion, it would make more sense to pursue this idea in the IDR working group. Margaret 2001 Message-Id: Date: Thu, 6 Dec 2001 12:59:49 -0500 From: Eric Gray Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Margaret, Working out how an implementation might distribute BGP computations within a single BGP node is definitely not a topic for the IDR working group. Whether or not it is a topic for any IETF working group is another issue. You wrote: > Hi Mareline, > > >(1) We cant have each router peering up with the other router to form a > >full IBGP mesh as this would lead to scalability issues. > > Are there any scalability issues introduced by IBGP that are not > inherent in distributing this information between a mesh of nodes using > any unicast communications channel? If so, you should probably raise > this issue with the IDR working group (responsible for BGP). > > >(2) We can have a central server to which all BGP learnt routes are sent. > >There it runs its decision process and there propagates those selected routes > >to every other router. This creates a single point of failure. Moreover, > >with the full BGP routing table being injected there is going to be a > >considerable load on the designated router/processor. > > How would this be different than running a single BGP on a single > control element and having the distributed nodes forward all BGP > traffic to that control element? In other words, if the BGP > protocol is not well-suited to being distributed across multiple > CEs, why distribute it? > > >(4) Hack the BGP protocol and making it TCP independent, broadcasting all > >the BGP packets within the consolidated router for within this integrated > >networking element all routers running BGP are peers. Reliability is at > >stake (which we don't ever bother with TCP!). > > A multicast version of IBGP is an interesting idea. It would probably > scale well when all IBGP nodes within a single group are attached to a > single ethernet (as in your example). How would it scale if the IBGP > nodes were distributed throughout a complex routing domain? > > >Is there any thought in the working group on standardizing such > >distributed BGP route computations? > > In my opinion, it would make more sense to pursue this idea in the > IDR working group. > > Margaret -- Eric Gray (mailto:eric.gray@sandburst.com) http://www.mindspring.com/~ewgray 2001 Message-Id: Date: Thu, 6 Dec 2001 12:54:24 +0530 From: rana pratap sircar Organization: Wipro Technologies MIME-Version: 1.0 Content-Type: multipart/mixed; boundary="------------5D383AD36040BA34243376D1" This is a multi-part message in MIME format. --------------5D383AD36040BA34243376D1 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Hi Mareline, The assumption here is that the two network elements are completely independent entities. However, in most of the implementations, there will be a system control card. This card can hold the information centrally. I think that this is more of a implementation issue than a standardization issue. Thanks & regards, Rana Mareline Sheldon wrote: > Hi All, > The mail is attached for it has a diagram which would be distorted > otherwise. > > regards, > mareline s > > ------------------------------------------------------------------------ > Name: bgp_distribution.txt > bgp_distribution.txt Type: Plain Text (text/plain) > Encoding: quoted-printable --------------5D383AD36040BA34243376D1 Content-Type: text/x-vcard; charset=us-ascii; name="rana.sircar.vcf" Content-Transfer-Encoding: 7bit Content-Description: Card for Rana Pratap Sircar Content-Disposition: attachment; filename="rana.sircar.vcf" begin:vcard n:Sircar;Rana Pratap tel;fax:+91-80-5732696 / 5732441 tel;home:+91-80-5278075 tel;work:+91-80-5732293 extn. 3029/ VoIP- 8034038 x-mozilla-html:FALSE url:www.wipro.com org:Telecom & Inter-Networking Solutions version:2.1 email;internet:rana.sircar@wipro.com title:Architect - Product Realisation adr;quoted-printable:;;Wipro Technologies =0D=0A26, Chamundi Complex, =0D=0AHosur main Road, =0D=0ABommanahalli;Bangalore;Karnataka;560 068;India x-mozilla-cpt:;25920 fn:Rana Pratap Sircar end:vcard --------------5D383AD36040BA34243376D1 Content-Type: text/plain; name="InterScan_Disclaimer.txt" Content-Transfer-Encoding: 7bit Content-Disposition: attachment; filename="InterScan_Disclaimer.txt" ------------------------------------------------------------------------------------------------------------------------- Information transmitted by this E-MAIL is proprietary to Wipro and/or its Customers and is intended for use only by the individual or entity to which it is addressed, and may contain information that is privileged, confidential or exempt from disclosure under applicable law. If you are not the intended recipient or it appears that this mail has been forwarded to you without proper authority, you are notified that any use or dissemination of this information in any manner is strictly prohibited. In such cases, please notify us immediately at mailto:mailadmin@wipro.com and delete this mail from your records. ---------------------------------------------------------------------------------------------------------------------- --------------5D383AD36040BA34243376D1-- 2001 Message-Id: Date: Thu, 6 Dec 2001 12:28:12 -0500 From: Jamal Hadi Salim Organization: Znyx Networks Subject: Re: BGP in a distributed environment Comments: To: Margaret Wasserman MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: base64 DQpJIGFtIG1pc3Npbmcgc29tZXRoaW5nIGhlcmU7IHdoeSBJRFI/IFdoeSBzaG91bGQgQkdQIGJl IGEgc3BlY2FpbCBjYXNlPw0KVGhpcyBpcyB3aHkgd2UgaGF2ZSBGb3JDRVMgdG8gYmVnaW4gd2l0 aC4gWW91IGNvdWxkIGhhdmUgYSBzaW5nbGUgQkdQIGNvbXB1dGUgDQpub2RlIHdoaWNoIGNvdWxk IHNlcnZlIG11bHRpcGxlIE5FcyBhcyBzL2hlIHN1Z2dlc3QgaW4gaGlzIDIpDQpSZW1vdmluZyB0 aGF0IHNpbmdsZSBwb2ludCBvZiBmYWlsdXJlIGlzIG91dCBvZiB0aGUgc2NvcGUgb2YgRm9yQ0VT LiBJZiANCndpdGhpbiBhIHJlbGlhYmxlIGVudmlyb25tZW50IChzdWNoIGFzIGxpa2VseSB3aXRo aW4gdGhlIHNjb3BlIGEgc2luZ2xlIGJveCksIA0KbXVsdGljYXN0IGZvciBkaXN0cmlidXRpbmcg dGhlIEJHUCBjb21wdXRlIHJlc3VsdHMgdG8gdGhlIE5FIGlzIGEgZmFpciB1c2Ugb2YgDQp0aGUg c3RpbGwgdG8gYmUgZGVmaW5lZCBGb3JDRVMgcHJvdG9jb2wuIFRDUC9TQ1RQIG9ubHkgY29tZXMg aW50byBwbGF5IHdoZW4gDQpnb2luZyBhY3Jvc3MgbXVsdGlwbGUgaG9wcy4gVGhpcywgaSB0aG91 Z2h0IHdhcyBjbGVhciBpbiB0aGUgcmVxdWlyZW1lbnRzIGRvYy4NCg0KanVzdCBteSAkLjAyIGNl bnRzIC5jYQ0KDQpjaGVlcnMsDQpqYW1hbA0KDQpPbiBUaHVyc2RheSAwNiBEZWNlbWJlciAyMDAx IDEyOjA1LCBNYXJnYXJldCBXYXNzZXJtYW4gd3JvdGU6DQo+IEhpIE1hcmVsaW5lLA0KPg0KPiAg PigxKSBXZSBjYW50IGhhdmUgZWFjaCByb3V0ZXIgcGVlcmluZyB1cCB3aXRoIHRoZSBvdGhlciBy b3V0ZXIgdG8gZm9ybSBhDQo+ICA+ZnVsbCBJQkdQIG1lc2ggYXMgdGhpcyB3b3VsZCBsZWFkIHRv IHNjYWxhYmlsaXR5IGlzc3Vlcy4NCj4NCj4gQXJlIHRoZXJlIGFueSBzY2FsYWJpbGl0eSBpc3N1 ZXMgaW50cm9kdWNlZCBieSBJQkdQIHRoYXQgYXJlIG5vdA0KPiBpbmhlcmVudCBpbiBkaXN0cmli dXRpbmcgdGhpcyBpbmZvcm1hdGlvbiBiZXR3ZWVuIGEgbWVzaCBvZiBub2RlcyB1c2luZw0KPiBh bnkgdW5pY2FzdCBjb21tdW5pY2F0aW9ucyBjaGFubmVsPyAgSWYgc28sIHlvdSBzaG91bGQgcHJv YmFibHkgcmFpc2UNCj4gdGhpcyBpc3N1ZSB3aXRoIHRoZSBJRFIgd29ya2luZyBncm91cCAocmVz cG9uc2libGUgZm9yIEJHUCkuDQo+DQo+ICA+KDIpIFdlIGNhbiBoYXZlIGEgY2VudHJhbCBzZXJ2 ZXIgdG8gd2hpY2ggYWxsIEJHUCBsZWFybnQgcm91dGVzIGFyZSBzZW50Lg0KPiAgPlRoZXJlIGl0 IHJ1bnMgaXRzIGRlY2lzaW9uIHByb2Nlc3MgYW5kIHRoZXJlIHByb3BhZ2F0ZXMgdGhvc2Ugc2Vs ZWN0ZWQNCj4gID4gcm91dGVzIHRvIGV2ZXJ5IG90aGVyIHJvdXRlci4gVGhpcyBjcmVhdGVzIGEg c2luZ2xlIHBvaW50IG9mIGZhaWx1cmUuDQo+ICA+IE1vcmVvdmVyLCB3aXRoIHRoZSBmdWxsIEJH UCByb3V0aW5nIHRhYmxlIGJlaW5nIGluamVjdGVkIHRoZXJlIGlzIGdvaW5nDQo+ICA+IHRvIGJl IGEgY29uc2lkZXJhYmxlIGxvYWQgb24gdGhlIGRlc2lnbmF0ZWQgcm91dGVyL3Byb2Nlc3Nvci4N Cj4NCj4gSG93IHdvdWxkIHRoaXMgYmUgZGlmZmVyZW50IHRoYW4gcnVubmluZyBhIHNpbmdsZSBC R1Agb24gYSBzaW5nbGUNCj4gY29udHJvbCBlbGVtZW50IGFuZCBoYXZpbmcgdGhlIGRpc3RyaWJ1 dGVkIG5vZGVzIGZvcndhcmQgYWxsIEJHUA0KPiB0cmFmZmljIHRvIHRoYXQgY29udHJvbCBlbGVt ZW50PyAgSW4gb3RoZXIgd29yZHMsIGlmIHRoZSBCR1ANCj4gcHJvdG9jb2wgaXMgbm90IHdlbGwt c3VpdGVkIHRvIGJlaW5nIGRpc3RyaWJ1dGVkIGFjcm9zcyBtdWx0aXBsZQ0KPiBDRXMsIHdoeSBk aXN0cmlidXRlIGl0Pw0KPg0KPiAgPig0KSBIYWNrIHRoZSBCR1AgcHJvdG9jb2wgYW5kIG1ha2lu ZyBpdCBUQ1AgaW5kZXBlbmRlbnQsIGJyb2FkY2FzdGluZyBhbGwNCj4gID50aGUgQkdQIHBhY2tl dHMgd2l0aGluIHRoZSBjb25zb2xpZGF0ZWQgcm91dGVyIGZvciB3aXRoaW4gdGhpcyBpbnRlZ3Jh dGVkDQo+ICA+bmV0d29ya2luZyBlbGVtZW50IGFsbCByb3V0ZXJzIHJ1bm5pbmcgQkdQIGFyZSBw ZWVycy4gUmVsaWFiaWxpdHkgaXMgYXQNCj4gID5zdGFrZSAod2hpY2ggd2UgZG9uJ3QgZXZlciBi b3RoZXIgd2l0aCBUQ1AhKS4NCj4NCj4gQSBtdWx0aWNhc3QgdmVyc2lvbiBvZiBJQkdQIGlzIGFu IGludGVyZXN0aW5nIGlkZWEuICBJdCB3b3VsZCBwcm9iYWJseQ0KPiBzY2FsZSB3ZWxsIHdoZW4g YWxsIElCR1Agbm9kZXMgd2l0aGluIGEgc2luZ2xlIGdyb3VwIGFyZSBhdHRhY2hlZCB0byBhDQo+ IHNpbmdsZSBldGhlcm5ldCAoYXMgaW4geW91ciBleGFtcGxlKS4gIEhvdyB3b3VsZCBpdCBzY2Fs ZSBpZiB0aGUgSUJHUA0KPiBub2RlcyB3ZXJlIGRpc3RyaWJ1dGVkIHRocm91Z2hvdXQgYSBjb21w bGV4IHJvdXRpbmcgZG9tYWluPw0KPg0KPiAgPklzIHRoZXJlIGFueSB0aG91Z2h0IGluIHRoZSB3 b3JraW5nIGdyb3VwIG9uIHN0YW5kYXJkaXppbmcgc3VjaA0KPiAgPmRpc3RyaWJ1dGVkIEJHUCBy b3V0ZSBjb21wdXRhdGlvbnM/DQo+DQo+IEluIG15IG9waW5pb24sIGl0IHdvdWxkIG1ha2UgbW9y ZSBzZW5zZSB0byBwdXJzdWUgdGhpcyBpZGVhIGluIHRoZQ0KPiBJRFIgd29ya2luZyBncm91cC4N Cj4NCj4NCj4gTWFyZ2FyZXQNCg== 2001 Message-Id: Date: Thu, 6 Dec 2001 12:05:20 -0500 From: Margaret Wasserman Subject: Re: BGP in a distributed environment Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Hi Mareline, >(1) We cant have each router peering up with the other router to form a >full IBGP mesh as this would lead to scalability issues. Are there any scalability issues introduced by IBGP that are not inherent in distributing this information between a mesh of nodes using any unicast communications channel? If so, you should probably raise this issue with the IDR working group (responsible for BGP). >(2) We can have a central server to which all BGP learnt routes are sent. >There it runs its decision process and there propagates those selected routes >to every other router. This creates a single point of failure. Moreover, >with the full BGP routing table being injected there is going to be a >considerable load on the designated router/processor. How would this be different than running a single BGP on a single control element and having the distributed nodes forward all BGP traffic to that control element? In other words, if the BGP protocol is not well-suited to being distributed across multiple CEs, why distribute it? >(4) Hack the BGP protocol and making it TCP independent, broadcasting all >the BGP packets within the consolidated router for within this integrated >networking element all routers running BGP are peers. Reliability is at >stake (which we don't ever bother with TCP!). A multicast version of IBGP is an interesting idea. It would probably scale well when all IBGP nodes within a single group are attached to a single ethernet (as in your example). How would it scale if the IBGP nodes were distributed throughout a complex routing domain? >Is there any thought in the working group on standardizing such >distributed BGP route computations? In my opinion, it would make more sense to pursue this idea in the IDR working group. Margaret 2001 Message-Id: Date: Thu, 6 Dec 2001 11:36:34 +0530 From: Mareline Sheldon MIME-Version: 1.0 Content-Type: multipart/mixed; boundary="----=_NextPart_000_0377_01C17E4A.42626970" This is a multi-part message in MIME format. ------=_NextPart_000_0377_01C17E4A.42626970 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 7bit Hi All, The mail is attached for it has a diagram which would be distorted otherwise. regards, mareline s ------=_NextPart_000_0377_01C17E4A.42626970 Content-Type: text/plain; name="bgp_distribution.txt" Content-Transfer-Encoding: quoted-printable Content-Disposition: attachment; filename="bgp_distribution.txt" Hi, A network element is composed of numerous logically separated entities = that cooperate to provide a given functionality (such as a routing or IP = switching) and yet appear as a normal integrated network element to = external entities. Two primary types of network element components = exist: control-plane components and forwarding-plane components (ASIC). = A standard set of mechanisms for connecting these components would = provide increased scalability and allow the control and forwarding = planes to evolve independently thus promoting faster innovation. Going a step further, we conglomerate these various network elements and = put them on an ethernet to form a truly integerated network element = [read as, terrabit router] composed of multiple packet processing = engines interconnected to form a logical whole. [Ref to the figure below] -------------------------------------------------------------------------= --------------- | -------------------------------- = ------------------------------- | | | NE1 | | NE2 = | | | | ------------------------------------------- = | | | | | | | | = | | | | ------------- | | ------------- = | | | | | CE | | | | CE | = | | | | ------------- | | ------------- = | | | | / \ | | / \ = | | | | / \ | | / \ = | | | | / \ | | / \ = | | | | / \ | | --------- = --------- | | | | ----------- ----------- | | | FE | | FE = | | | | | | FE | | FE | | | ---------- = --------- | | | | ----------- ----------- | | | | | | | = | | | | | | | | | | | | | | | | | | | | = | | | | | | | | | | | | | | | | | | | | = | | | | | | | | | | | | | | = -------------------------------- | | -------------------------------- | | | | | = | | | | | | | | | | | | | | | | = | | -------------------------------------------------------------------------= --------------- | | | | | | | | | | | | | | ^ ^ | | BGP BGP NE Network element CE Control element=20 FE Forwarding element Suppose we are running BGP on one of the interfaces of one network = element [aka router]. Whatever routes it learns must be advertised to = all the remaining network elements. This can be done via some = redistribution mechanism within the router. The problem comes in when we = are running BGP on another interface belonging to some other NE (Say 2 = in our case). Now whatever routes NE1 learns, must be passed along to = NE2 with *ALL* its attributes so that it can run its decision process = and come up with the optimal routes. The same holds true for NE1. It = must have all the routes along with the complete attributes from NE2 so = that it can run its decision process. Once they have a consistent table, = the routes can be leaked into the remaining network elements [routers]. = This problem aggravates when we have network elements [routers] of the = order of O(1000) What are the options left with us to ensure that the BGP tables across = the NEs remain synchronized. (1) We cant have each router peering up with the other router to form a = full IBGP mesh as this would lead to scalability issues. (2) We can have a central server to which all BGP learnt routes are = sent. There is runs its decision process and there propagates those = selected routes to every other router. This creates a single point of = failure. Moreover, with the full BGP routing table being injected there = is going to be a considerable load on the designated router/processor. (3) We can have hierarchical route reflectors within the router but it = is also not scalable. Secondly, there would be a lot of complexities = involved in the selection of server and the clients, etc. (4) Hack the BGP protocol and making it TCP independent, broadcasting = all the BGP packets within the consolidated router for within this = integrated networking element all routers running BGP are peers. = Reliability is at stake (which we dont ever bother with TCP!).=20 Is there any thought in the IDR working group on standardizing such = distributed BGP route computations ? Regards, Manav Bhatia ------=_NextPart_000_0377_01C17E4A.42626970-- _________________________________________________________ Do You Yahoo!? Get your free @yahoo.com address at http://mail.yahoo.com 2001 Message-Id: Date: Thu, 6 Dec 2001 11:34:32 -0800 From: "Anderson, Todd A" Subject: Re: BGP in a distributed environment MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable People have brought up this idea before of having different routing protocols running on different CEs but the problem comes when those protocols want to apply what they have learned. Essentially, you need some arbiter that decides which protocol's routes take precedence. I don't think you want FEs to have the sort of knowledge necessary to make this decision. My view is that you need a centralized route arbiter in one of the CEs to which all protocols send their results and the arbiter produces a unified table that it then sends diffs to the FEs. However, how the CEs communicate with each other or the arbiter I think is currently out of scope of ForCES and the = requirements document places no restrictions on CE to CE communication. Todd -----Original Message----- From: Martin, Olivier [mailto:Olivier.Martin@versalys.com] Sent: Thursday, December 06, 2001 10:32 AM To: FORCES@PEACH.EASE.LSOFT.COM Subject: Re: [FORCES] BGP in a distributed environment Hi everyone, I might be mistaking, but the same issue would arise with any IGP too. There could be a different central CE per routing protocol instance, that would send the forwarding tables along with the results of its calculation to the FEs, which in turn would use that information to populate their own forwarding tables in order to properly send the egress traffic to the proper interface / FE. The same information should be sent out to other CEs in the system, so if there are dependencies between some IP service and another one (such as the relationship between BGP and the IGP within the AS), the CEs would be aware of it. All of this communication should be using the mecanism described in the requirement draft, over a the [RFC2914]-compliant protocol. =20 This would create the possibility of having N CEs, each taking care of = a particular function within the NE and since these entities are using GP CPUs, they could take over the process of some failed CE, using some VRRP or similar functionnality. =20 My 2 cents.ca too ;) !=20 Cheers, Olivier -----Message d'origine----- De : Margaret Wasserman [mailto:mrw@windriver.com] Envoy=E9 : 6 d=E9cembre, 2001 12:05 =C0 : FORCES@PEACH.EASE.LSOFT.COM Objet : Re: BGP in a distributed environment Hi Mareline, >(1) We cant have each router peering up with the other router to form a >full IBGP mesh as this would lead to scalability issues. Are there any scalability issues introduced by IBGP that are not inherent in distributing this information between a mesh of nodes using any unicast communications channel? If so, you should probably raise this issue with the IDR working group (responsible for BGP). >(2) We can have a central server to which all BGP learnt routes are sent. >There it runs its decision process and there propagates those = selected routes >to every other router. This creates a single point of failure. Moreover, >with the full BGP routing table being injected there is going to be a >considerable load on the designated router/processor. How would this be different than running a single BGP on a single control element and having the distributed nodes forward all BGP traffic to that control element? In other words, if the BGP protocol is not well-suited to being distributed across multiple CEs, why distribute it? >(4) Hack the BGP protocol and making it TCP independent, broadcasting all >the BGP packets within the consolidated router for within this integrated >networking element all routers running BGP are peers. Reliability is at >stake (which we don't ever bother with TCP!). A multicast version of IBGP is an interesting idea. It would probably scale well when all IBGP nodes within a single group are attached to a single ethernet (as in your example). How would it scale if the IBGP nodes were distributed throughout a complex routing domain? >Is there any thought in the working group on standardizing such >distributed BGP route computations? In my opinion, it would make more sense to pursue this idea in the IDR working group. Margaret 2001 Message-Id: Date: Thu, 6 Dec 2001 10:56:46 -0800 From: Alex Zinin Organization: Nexsi Systems Subject: Re: BGP in a distributed environment Comments: To: Mareline Sheldon MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Mareline, > Is there any thought in the working group on standardizing such > distributed BGP route computations? Yes, there is a thought: please try to stay away from specifying how things should be implemented. -- Alex Zinin Thursday, December 06, 2001, 2:18:18 AM, Mareline Sheldon wrote: > Hi All, > The mail is attached for it has a diagram which would be distorted > otherwise. > regards, > mareline s 2001 Message-Id: Date: Tue, 4 Dec 2001 09:44:13 -0800 From: "Putzolu, David" Subject: Draft Agenda for ForCES at IETF 52 MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" All, Below is the agenda that Patrick and I have come up with. For any changes please email Patrick and myself asap as the deadline for changes is Wednesday (tomorrow) at 9:00am PST. Cheers, David --- cut here --- IETF 52 Forwarding and Control Element Separation WG (ForCES) Agenda Monday, December 10 at 1930-2000 ================================ CHAIRS: Patrick Droz David Putzolu AGENDA: 1930 Administrative Matters (chairs, 10) - Selection of minute-taker - Circulation of attendance sheets - Agenda bashing 1940 Requirements for Separation of IP Control and Forwarding (Doria, 30) http://www.ietf.org/internet-drafts/draft-ietf-forces-requirements-01.txt 2010 ForCES Applicability Statement (Handley, 20) http://www.ietf.org/internet-drafts/draft-crouch-forces-applicability-00.txt 2030 ForCES Architectural Framework and FE Functional Model (Anderson, 30) http://www.ietf.org/internet-drafts/draft-anderson-forces-model-00.txt 2100 Wrap-up 2001 Message-Id: Date: Mon, 3 Dec 2001 18:08:54 -0500 From: "Gopal Ram (NRC/Boston)" Subject: Question in applicability and Requirement document. MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Hello All, Following are the question which i need more information. =20 Question (1)=20 Document Name: Forces Applicability Statements Refer Section 3.1.1 Discovery Is this within the scope , this contradicts with requirement document=20 where it is stated that pre association phase ( in definition section )is out of scope.=20 If this is out of scope we do not need this section in this document. =20 Question (2)=20 Document Name: Forces Requirement =20 Refer : Section 2, Definition=20 No where else these terms like FE Manager or CE Managers are used in the document. It would be better if you could add a figure or some sentence regarding the placement where there placement are. IT is not clear to me where is to be placed in the network. =20 Question (3)=20 Document Name: Forces Applicability Statements Refer : Section 3.3 CE/FE Locality. [1] If the CE and FE are not the single physical unit , whether it is one hop away or multiple hops away does not make a difference ( of course it should be within a LAN under single administration) [2] If CE is away from FE , how it is different if we configure thro' normal SNMP/CLI etc or other=20 configuration procedure. What is the advantage of using this? Regards Ramg =20 2001 Message-Id: Date: Mon, 3 Dec 2001 16:16:22 -0800 From: "Anderson, Todd A" Subject: Re: Question in applicability and Requirement document. MIME-Version: 1.0 Content-Type: text/plain; charset="ISO-8859-1" My comments inline. Todd > -----Original Message----- > From: Gopal Ram (NRC/Boston) [mailto:ram.gopal@nokia.com] > Sent: Monday, December 03, 2001 3:09 PM > To: FORCES@PEACH.EASE.LSOFT.COM > Subject: [FORCES] Question in applicability and Requirement document. > > > Hello All, > Following are the question which i need more information. > > Question (1) > Document Name: Forces Applicability Statements > Refer Section 3.1.1 Discovery > > Is this within the scope , this contradicts with requirement document > where it is stated that pre association phase ( in definition section > )is out of scope. > > If this is out of scope we do not need this section in this document. I believe that the intent was to say that we believe discovery is out of scope in both documents. The applicability statement does not explicitly say the words "out of scope" but it does say that the process called discovery has already happened before what we are calling ForCES has started. If the documents were silent on the discovery issue then people would either assume that it was part of the ForCES protocol or think we were missing a piece of the puzzle. I think that the purpose of both documents in this respect is to say that we realize that the discovery process needs to somehow happen but that to make progress we want are currently defining discovery as out of scope. > Question (2) > Document Name: Forces Requirement > Refer : Section 2, Definition > > No where else these terms like FE Manager or CE Managers are used in > the document. > It would be better if you could add a figure or some > sentence regarding > the placement > where there placement are. IT is not clear to me where is to > be placed > in the network. This relates to your first question. We defined the CE manager, FE manager, and pre-association phase protocol for the purpose of saying that the discovery process was currently out of scope and to say that discovery was not part of the ForCES protocol. Since we defined discovery as out of scope, it is not surprising that in the rest of the document we place no requirements on that process or any of the entities involved. To answer the second part of your question, the wording in the definition is meant to say that you can combine FE managers and FEs or you can have them separate. As a personal opinion, I think there will be two common cases. The first is a "DHCP" FE manager that is physically integrated with a FE and uses DHCP or a DHCP-like protocol to find the closest CE. The second case is one where there is a single FE manager integrated with a management console and that FE manager is responsible for a whole domain of FEs. The FE manager would tell each FE to which CE to communicate. The short answer to your question is though, you can put CE managers or FE managers anywhere you like. > Question (3) > Document Name: Forces Applicability Statements > Refer : Section 3.3 CE/FE Locality. > > [1] If the CE and FE are not the single physical unit , whether it is > one hop away or multiple hops away does not make a difference If it is only a single hop, you could implement ForCES directly over layer-2. If it is multi-hop then our charter mandates that we use a TCP-friendly transport protocol. The approach the design team agreed to was that we would define what the ForCES protocol expected of its underlying layer and that the first thing we would do is to define ForCES over a TCP- friendly layer-4 transport protocol. We saw this approach as leaving room in the future for people to define other encapsulations for single hop connections. > ( of course it should be within a LAN under single administration ) The "single administration" restriction is not one that the requirements design team made. Given that we are defining an IP protocol, it would be impossible a priori (at least for IPv4) to know whether a CE wanting to control a FE was in the same administrative domain or not. Even within the same administrative domain, you still need security so that unauthorized CE's can't start controlling FEs. The security considerations section of the draft addresses exactly this issue. Having said that, some people I've talked to have seen a market for a FE operator to lease control of one or more of its FEs to CEs outside the operator's administrative domain. I'm not sure there would be any way to prohibit this even if we wanted to and the requirements draft therefore says nothing about this issue. > [2] If CE is away from FE , how it is different if we configure thro' > normal SNMP/CLI etc or other > configuration procedure. What is the advantage of using this? By CLI do you mean humans doing the configuring? The interface between CE and FE has to be automated. Let's say your routing protocol in the CE learned a new route. You wouldn't want to wait for a human to install the corresponding forwarding entry on a FE. Am I understanding your question right? In terms of SNMP, we haven't started discussing protocols and therefore we have not ruled out any protocols at this point. However, as far as I know, SNMP runs over UDP and therefore fails ForCES protocol requirement #4. Essentially the problem is that the control timescale is much shorter than the management timescale and may takes up enough bandwidth that congestion control becomes an issue.