TSVAREA, IETF-73, Minneapolis, MN, USA WEDNESDAY, November 19, 2008 Chairs: Magnus Westerlund Lars Eggert Notes: Gorry Fairhurst (THANK YOU) * Review of Area (Lars and Magnus) PCN finished ARCH document The ledbat working group had been approved (from TANA BOF) and will be meeting this IETF meeting. IPPM and nfsv4 are rechartering. There will also be a discussion of the Router Alert option in the routing area this IETF meeting, that will have transport implications (e.g. NSIS, RSVP). * PCN WG Status (Philip Eardley) Presented a summary of new and completed work of the PCN WG, no questions. * Breaking Up the Transport Logjam (Bryan Ford) Stuart Cheshire: Appletalk did this (port numbers in datagram header). Bryan Ford: Yes, there are other protocols too. Joe Touch: For exactly the same reason, protocol numbers are examined, SYN fields of TCP are examined, and also things in the app stream. Why do we have to go to this extreme? Bryan Ford: Presentation will show this. Joe Touch: The DNS helps you find services (e.g. not caring of transport). Why do we not leave this as the service level? Andrew McGregor: Endpoint may not be in the DNS. Joe Touch: Could be using Bonjour. Bryan Ford: We could take the view that DNS as the place for discovery, or we could do something before DNS. Hannes Tschofenig: The DIME work uses both TCP and SCTP (we use DNS and manual configuration). Bryan Ford: The talk is about proposing a new architectural view. Matt Mathis: I think this is good, and the multiple transport connection could be implemented relatively easily. Randy Stewart: I disagree that you can not deploy protocols, just because they use NATs. SCTP is widely deployed. I think transports need to be in system software. Stanislav Shalunov: This is the Internet engineering task force - most deployed hosts are behind NATs. You also need to legacy issues. Bryan Ford: Yes, new stuff needs to work with old, there's more detail. Gorry Fairhurst: Simply adding a UDP header does not solve this. If you need a new transport, why grow these. Bob Briscoe: There is an issue of difference between an attachment point and other points need to be separated. There are two side that control the attachment points. Bryan Ford: This is linked to identity/locator separation. Remi Denis-Courmont: I am not sure you can separate this for congestion control. Michael Tuexen: Where do you think the split is to be performed in the kernel or user space? Bryan Ford: I don't care. Michael Tuexen: Congestion control should not be in user land. Bob Briscoe: I depart from you at the flow level. Bryan Ford: The idea is to separate the action of the congestion control loops. The end to end flow rate has. Bob Briscoe: This seems to be making this complicated, the main thing is to get the flow correct end-to-end. Carsten Bormann: Where do we put IPsec in this architecture? Bryan Ford: I do not know yet. Stanislav Shalunov: I like the addressing part, but the idea of multiple new routers that do many things seems not like the Internet. Bryan Ford: This is really split-TCP with an additional transport layer. Chris ?: This seems to be standardising what we have. Is this really turning the end-to-end principle on its head? Bryan Ford: No. * UDP-Encapsulated Transport Protocols (Remi Denis-Courmont) Stuart Cheshire: All firewalls do not stop SYN/SYNACK; all firewalls pass UDP. Remi Denis-Courmont: I am not saying that, it is more complex. Tim Shephard: There are protocols that need port information (e.g. DNS). Remi Denis-Courmont: We do not need DNS NAT to NAT, we could use SDP. Gorry Fairhurst: Why can you not also do this for UDP-Lite - if you care about getting things end to end for deployment, and need to do this, we should do this for all transports (just to let them be deployed). Remi Denis-Courmont: Not sure how we can get benefit if we encapsulate this end-to-end? Michael Tuexen: There are also issues with the Mux tag in SCTP. * Requirements for Pseudowire Congestion Control (David Black) Presentation was first made in PWE. Joe Touch: This is vaguely reminiscent of what I said in a discussion on BGP/TCP. If the TCP conenction fails, the BGP route is torn down. David Black: This is something similar, we have few options, and switching of the network path is one option that is on the table. Joe Touch: I would call that "path-collapse detection". Matt Mathis: I would prefer "overload protection". * Rethinking the "TCP-Friendly" Paradigm (Matt Mathis) Presentation was first made in IRTF ICCRG. Bob Briscoe: I think resource-sharing is a network problem. I think we need to think about failures in the Internet world. Bryan Ford: I agree with this. Stanislav Shalunov: I am glad to hear the change. Congestion control is like insurance. There is insurance against failure (that belongs to the net) there is also medical insurance (this type of day-to-day things belongs in the application) Bob Briscoe: I agree with Stas. Iljitsch van Beijnum: Not all loss is congestion. * Use of the IPv6 Flow Label as a Transport Layer Nonce (Steven Blake) Presentation was first made in 6man. Bob Briscoe: That is nice. I'd like to state it can also be used in ESP and hidden from the network. Iljitsch van Beijnum: Is this a system-wide change? Steven Blake: System-wide. Iljitsch van Beijnum: I Have another use of flow-label. Please do not enable this for all applications. Do not do for stuff that does not need it. Erik Nordmark: There is support in the API. Do both ends have to remember the flow-id for two MSL. Steven Blake: Only the one with time-waut state. Joe Touch: Declare 0 to mean that this is not used. Steven Blake: It is so. Joe Touch: There are also issues with NATs when both ends start at the same time. Fernando Gont: There is an algorithm in the port-randomisation draft that says how new sequence numbers can be chosen.