IETF85 - IRS BOF

// Etherpad is Full -- only 16 people are allowed to edit

Administrativia:

Ross Callon & Dave Ward:

• NOTE WELL et al.

Adrian:

• Working Group-forming BOF, meaning will investigate WG formation
• Persuading the IESG that there is scope for WG -> work to be done.

IRS -- Framework documents:

Joel Halpern:

• 4 I-Ds included
• Application that want to dynamically interact with the Internet (need filter, need policy, etc)
• Need models: what is the network, and what do I want to tell the network to do. // SNMP is not a good one. How to model policy?
• Applications integrate *into* the network
• Protocols are down the thinking roadmap. Focus: data models, security, authorization & identity.
• Programmatic interface async and fast.
• IRS complements -- it is not only config or replaces NMS/routing/signaling
• Start -- data-model RIB
• Scope (I can read) vs. Influence (can change)
• EOP (End of Presentation)
• Q at the Mike:
• Q Kireeti: what is an application? Netflix or topology manager (tied to the network)?
• A Joel: Real answer is both -- first will be operation control
• Q Kireeti: Policy -- what policy? 
• A Joel: Policy doc is policy for IRS. This is "IRS Policy". Not good terminology
• Q Kireeti: IRS is not a direct replacement for routing protocols. Is it indirect replacement?
• A Joel: Some people think we could replace in some deployments
• Q Vijay G.: Paris I2AEX BOF.
• A Joel: I was not there. Send the summary to the list
• Q Bumit B.: Do not use "IRS" use another acronym. Is it a network (system) API or a router (device) API?
• A Joel: These are routers now but could be distributed systems (if FORCES). 
• Q B: Are you aware of similar work on other SDOs?
• Q Tom Nadeau channeling someone from Cisco: IRS augments not replaces routing
• Q someone: Assuming multiple controllers talking to the device? Finite number of resources. How orchestrate decision?
• A: Discussed in the policy document -- needs to be addressed.
• Q someone: The term Application is confusing to people. Comment: now you can replace routing protocols the way it is written whether you intend to or not.

IRS - Use Case & Requirements:

Shane Amante:

• List of 5 drafts with use cases -- plus one more with narrowly defined use cases for Optical
• Caveat: Not intending that IRS needs to solve for all of these & read all drafts
• Amante-topology-usecases -- topology, policy & orchestration "Manager" functions
• There are already "analytics" systems that the manager would need to query to get link until, etc.
• Keyur-IRS-BGP-Usecases -- interesting. BGP config is likely the largest config in the box and complex. So this is useful. Also MPLS-TE.
• Russ-W.-IRS-use-case: overall fine tuning traffic flows in IP network. Dynamically optimizing in H&S.
• Medved-irs-topo-requirements: We do not have agreed on terminology. This is northbound API. This is from Commissioner northbound to the apps above it. Next I-D is southbound from IRS commissioner to IRS agent. pub-sub, flow-control, etc.
• rfernando-irs-framework-requirement: commissioner towards client (southbound). Very good draft, lots of detail, pub-sub included. Initially focus on p2p with reliable in-order delivery. App. above commissioner should not worry about how to talk o lower layers
• Open Mike:
• Russ White: This seems like boiling the ocean. Choosing which part of the ocean to boil. But we are trying to find out what the ocean looks like. 
• David: Two example use cases. If NETMOD finishes router config interface, when do I use which one?
• Shane: Last mike of IP renumbering or AS renumbering. Other work to be done for policy. You really need to have the notion of policy. One potential avenue is IRS uses NETCONF
• xxx Infinera: Nice cases but you start with solution, and then describe use cases.
• Shane: important to recognize that there are existing systems in mid-large networks (data collection warehouses) That's why we are taking this route. In the future we should get it out of there and incorporate it into the official framework
• xxx: Are you looking at actual topo of optical layer?
• Shane: Trying to propose a narrowly-defined use case that is extremely useful. I would like to get that full view, but I'd rather we scope narrow.
• Are you trying to define OTN overlays? SNMP exists.It is confusing to me to see IRS will provide things that are there.
• Shane: I agree -- but not my I-D.
• Will IRS provide an interface for the data warehouse? That's dangerous
• Shane: I want things that will make it to real networks. I want an API on top of those warehouses, any schema (JSON, XML). 
• This is very general. Why don't you use SQL?
• Adrian: Not charter to write another protocol.
• xxx DT: Confused… Scope? Ocean boiling?
• Shane: This is entirely taxonomy of use cases. This does not mean we need to solve all

Moving towards an IRS WG Charter:

Ross Callon:

• Observation about charters: focus, limited, well defined. And no overlap.
• Sent early charter to irs-discuss, with these issues:
• Allergic reactions to "slow vs. fast path", and I'm sympathetic.
• don't confuse implementation vs. standard.
• Which Protocol will IRS use to set or read? New or existing (with or without tweaks)? Will need to figure it out. Charter says no change to protocols (new or extend).
• Use cases: tightly scoped! preclude others? no, but let's focus.
• Charter: "IRS", let's change it so as not to alias with this American groups that comes after you if you do not fill certain forms. Need Operations Area advisor.
• EOF -- Questions?
• Ross: actually next preso is on obvious questions.

Obvious Questions:

Dave Ward:

• How big is IRS? 408 members on list. Operators, vendors, and academics.
• Use Cases: At least one use case, and no upper limit. BUT evaluation of use cases is examples and not canonical designs.
• My encoding is the prettiest. No, it isn't. None perfect today. Start with data model.
• What about NetConf/Yang.
• They are candidates, but address requirements
• Advantages for using Netconf/Yang. (Yang modeling, Netconf transport), but Gaps: operational state semantics, missing features, etc.
• Team of 3 folks investigating applicability.
• IRS & Programmability.
• Concerns: Transport Protocols, Identity, many more…
• EOF -- Question on Charter?
• Lou Berger: Current definition of topology is broad. Topology requirement with L2, Optical, MPLS, etc. What's in an out according to the charter? I cannot tell.
• Ross: Do we decide this before the WG or by the WG?
• Lou: I think it needs to be part of the charter -- to narrow the scope.
• Margaret W.: Two concerns with this work:
• Serious & blocking: we tried to make protocols that explain how to manage devices, and problem is "if we can manage a subset, then people will manage other things" (proprietary CLI). If they have Python to the CLI, why use this? We are setting ourselves up for this in a config protocol. Why is this different?
• Scared when people tell so many times that they will not boil the ocean.
• Ross: Use cases is key to both your questions. This is only useful if there are use cases of significant value to operators that see the point in using it, but the amount of stuff in the router is small.
• Margaret: They will use other stuff for non-use case, why not use other stuff for everything?
• Ross: Answers from operators
• DT: This is not system-wide management. This does not define or implement system-wide mgmt.
• DWard: You mean "the network and not a router, right?" Agent: speaks protocol and decomposes message. Not the OS for the networking node. For example, speaks NETCONF and parses YANG
• Q: Is this the same as SDN
• DWard: Since SDN has emerged to be everything, it is.
• DWard: First define what we want out of the network. Joel has architectural picture.
• Shane Amante responding to Margaret. I am not concerned with rate of adoption. We need something that can evolve quickly because it is software. 
• Dmitry: Model needs information. Why are there limits on information nodes and not the common functions where this information is used. Why don't you start thinking about state models? 
• DWard: We can add other models. Good point. It is required.
• Dmitry: Agent is function with memory. Using the word "agent" implies understanding.
• Brian Dixon: slow vs. fast path -- vendors should not implement this as gateway to CLI. CLI could be commissioner within the box. To Margeret's concern: not protocol but functional integration. Like "manage the BGP control plane across the network managed here, not all functions."
• Wes: I do not think you are trying to boil the ocean. You are trying to boil another planet. The body of work you talk about is as big as the IETF.Just managing RIB is huge.
• Key issue is data modeling. Answer to Margaret: Before NETCONF, SNMP modeling was not convenient.
• DWard: focusing heavily on requirements and use cases, focusing on RIB, policy, and topology.
• Peter: Glad to see use case for optical systems, it's my problem.

Finale by DWard:

Show of hands of people interested in working on this work. (around over 1/2 the room raises their hands.)

Thank you very much

Show of hands of who believe this should become a chartered WG. (around over 1/2 the room raises their hands.)

Thank you very much

Show of hands of who believe this should NOT become a chartered WG. (3 hands)

Ross: AD, additional questions?

Adrian: Yes, I am nervous. Discussion says "nail the charter of small use cases", but everyone is thinking of their own use case.

DWard: Is this enough to charter the work? few hands

DWard: People think use cases should be fewer? 1/2 the room

DWard: Who believes that a whole different set of documents should charter? No hands.

Luca Martini: No document on security.

DWard: Multicast?

Adrian: Will we do IRS-TP?

David Ward: Yes, not in this standards body.

Adrian: Do people understand difference between information model and data model? RFC 3444

Kireeti: IRS MIB?

David: Thanks for volunteering...

Adrian: continue on the list.

David Ward: Thanks, see you at next IETF

Closure.