Congestion and Pre-Congestion Notification WG (pcn) IETF73 Meeting Minutes CHAIRs: Scott Bradner Steven Blake Minute taker: Tom Taylor TUESDAY, November 18, 2008 0900-1130 Morning Session I Salon E ==================================== AGENDA: o Administrivia Steven Blake presented the agenda slides: http://www3.ietf.org/proceedings/08nov/slides/pcn-0.pdf He reviewed the status of the charter milestones. The architecture document is green, most of the required documents are yellow, while the signaling requirements document is red. ------------------ o PCN Architecture draft-ietf-pcn-architecture-08 Philip Eardley discussed the status of the PCN architecture draft. Slides: http://www3.ietf.org/proceedings/08nov/slides/pcn-1.pdf The draft has been through two working group last calls. Currently waiting on Ad direction before submitting the draft for IESG approval (and IETF Last Call, according to Lars Eggert). He mentioned that volunteers were needed to write edge node behavior drafts. There will be a corridor meeting on Thursday to work on a template for these drafts. ------------------ o Marking Behavior draft-ietf-pcn-marking-behavior-01 Philip Eardley discussed the status of the PCN Marking Behavior draft. Slides: http://www3.ietf.org/proceedings/08nov/slides/pcn-2.pdf Philadelphia consensus on marking behavior was confirmed in Dublin, and documented here. Traffic conditioning section was rewritten. There is no normative advice on the prioritization of competing non-PCN packets. Philip considers the draft ready for working group last call. Chair asked how many in room had read document, stated that the number had to increase, and encouraged participants to read the draft and comment on the list. Philip mentioned a number of points which have had list discussion: o Satoh draft (draft-satoh-pcn-ST-marking-00.txt) o 1/N marking -- Michael Menth showed equivalent behavior could be achieved by performing 1/N reduction in edge termination. o Make preferential dropping optional rather than SHOULD; add option for random dropping (e.g. tail drop) Philip reviewed slides of different dropping policies. The effect of each depends on the marking behaviour. Michael Menth had suggested to make dropping behaviour a configurable option. This choice has pros and cons. Problem of what to say in a standards document. Suggests keeping the SHOULD on preferential dropping of marked packets, but add explanation of when it would be reasonable to use another behaviour (this was the outcome of a group discussion yesterday) > Ruediger Geib: too many options - gut feeling > Glen Zorn: IETF should not standardize configuration. And this doesn't affect the protocol, so why mention it? > Scott Bradner: does have operational significance. Does change behaviour; have to ensure interoperability. > Steven Blake: could set up to require configurability. Simplest way is to leave it up to operator to turn on a policer, turn on a marker. Should take discussion to list. > (Unknown): Bad idea to add configuration option. > (Unknown): Need to think carefully about describing behaviour that is not normative. SHOULDs should lead to interoperability. > Stu Goldman: Too many options is not a standard. Philip's closing remark -- 100% consensus is not achievable. Chairs agree, working group is aiming for rough consensus. The draft needs another round on the mailing list to close discussion points, before initiating a working group last call. ------------------- o Baseline Encoding draft-ietf-pcn-baseline-encoding-01 Bob Briscoe discussed the status of the Baseline Encoding draft. Slides: http://www3.ietf.org/proceedings/08nov/slides/pcn-3.pdf The draft has been updated twice since the Dublin meeting, and is now a working group draft. Main change to encoding is to specify that ECN='01' encoding is for Experimental (not local) use. Added a section on restrictions for extension encoding to ensure backward compatibility. Node behaviours moved to an appendix. Added change of terminology to 'PCN-compatible DSCP', 'PCN-enabled packet'. Summary slide at end shows why the baseline encoding is the best choice for extensibility. Open issue: where to put text on valid and invalid codepoint transitions? Other than this, Bob considers the draft ready for working group last call. Chair asked who had read this draft (fewer than for the marking behaviour draft). He encouraged participants to read it and comment on the mailing list. > Philip Eardley: people should especially review the section on restrictions for future extensions. > Lars Eggert (AD): Informational documents don't normally go through IETF Last Call, but lately architecture and requirements documents do. Approval of the architecture draft will be needed before other PCN drafts can progress. Chairs agreed that a request for publication of the architecture draft will be submitted shortly (Scott Bradner will be document shephard). AD review will be short because Lars has already read the draft. --------------------- o 3-in-1 PCN Encoding draft-briscoe-pcn-3-in-1-encoding-00 Bob Briscoe discussed the status of his 3-in-1 PCN Encoding draft. Slides: http://www3.ietf.org/proceedings/08nov/slides/pcn-4.pdf He is looking to developments elsewhere to enable this option. Expected disposition, all going well, is that it become an Experimental RFC. Need more people to read and comment on the PCN mailing list. -------------------- o PSDM Edge Behavior draft-menth-pcn-psdm-deployment-00 Philip Eardley discussed the status of the Michael Menth's PSDM Edge Behavior draft. Slides: http://www3.ietf.org/proceedings/08nov/slides/pcn-5.pdf Two aspects to the draft: encoding (experimental extension of baseline), plus a deployment model. Chair asked if the encoding used in this draft is compatible with work Bob Briscoe is doing, answer is yes. The slides review how PSDM works. Data and probe packets are marked differently. Admission control is performed on probe packets; flow termination is performed on data packets. An example shows RSVP PATH message as an implicit probe. A second example has explicit probe packets, regular probing, state transitions between admit flows, and do not admit flows. Slide on benefits. Note: the scheme works with multipath and small ingress-egress aggregates. > Bob Briscoe: one downside -- the fact that the type of marking is under the sender's control may open up an opportunity for abuse in the inter-domain case. ---------------------------------------------------------- o Single PCN-Threshold-Marking Using PCN Baseline Encoding draft-satoh-pcn-ST-marking-00 Daisuke Satoh discussed his draft on PCN single threshold marking. Slides: http://www3.ietf.org/proceedings/08nov/slides/pcn-6.pdf The scheme uses the baseline encoding, and provides explicit detection of flows exceeding the admission threshold and supportable rate. It uses a ratio marking in the state between admissable and supportable rate. Marking ratio is based on the degree of fill of the marking token bicket. Note that the marking ratio is cumulative over the successive links in the path -- longer paths imply higher marking rates, all else being equal. Use of a second threshold adds a break in the graph of marking rate vs. arrival rate. Marking on/off behaviour adds a second break. Satoh also discussed the edge behaviour for flow termination. > Bob Briscoe: assumption that all markinggs from one queue are associated to the same ingress. Satoh assured that all packets are marked in flow termination state. Concluded that there is no issue. The termination algorithm uses an additional term with factor y to ensure that termination is brought down to the supportable rate, not just link capacity. > Anna Charny: as agreed in discussion Monday, admission will be terminated even when operating below admission threshold if flows are bursty. Suggests that the time it takes to terminate should be a criterion for assessment of the algorithm. Choice of y to balance between over-termination and over- slow termination is difficult. > Bob Briscoe: question on motivation. Are you motivated by lack of codepoints, or is the scheme seen as valuable on its own merits? > Philip Eardley: rephrasing the question -- with additional codepoints, the scheme could see significant change. Satoh has no view at moment of what it would look like if the coding scheme allowed decoupling of admission and termination behaviours. > Bob Briscoe: more interested in whether they would use single marking at all if more codepoints available. > Anna Charny: disagree with this formulation. Satoh agreed that performance would improve with 3-state encoding. > Bob Briscoe: will the algorithm would be simplified? > Anna Charny: the algorithm might still stand on its own merits; should not automatically assume single marking-type algorithms are less desirable and should be set aside if codepoints allow alternatives. Chair requested further discussion of this point on list. > Philip Eardley: is further behaviour required? > Steven Blake: believe that this is the whole package. > Bob Briscoe: can get synchronization behaviour in bucket where the same packets for the same flow tend to arrive when the bucket is in the same state. Get biased handling of flows as a result. > Matt Matthis: clarified, picked-upon flows have to go to the same egress. Satoh mentioned that there is an element of randomization. > Bob Briscoe: but this assumes random traffic -- no randomization in algorithm. > Matt Mathis: in extreme case, the same flow would always be the one marked. > Anna Charny: is a more detailed evaluation available? Satoh said yes for admission behavior, not yet for termination behavior. > Anna Charny: summing up -- a number of perfomance concerns, need more info before changing what we have done so far. Chair requested futher discussion on the PCN mailing list. --------------- o ECN Tunneling Briscoe 10 min draft-ietf-tsvwg-ecn-tunnel-01 Bob Briscoe presented his TSVWG draft 'Layered encapsulation of congestion notification'. Slides: http://www3.ietf.org/proceedings/08nov/slides/pcn-7.pdf This draft proposes a change of behaviour of all ECN tunneling. Now a TSVWG draft in TSVWG. Question to the PCN WG is whether to add to our scope. Draft currently brings non-IPsec tunneling behaviour into line with IPsec behaviour. Proposed further change of behaviour at egress is currently in an appendix. Do we need to make this normative? Do we do it in the same draft? > Matt Mathis: are there considerations of deployment costs? > Bob Briscoe: can't think of any. Can make much shorter argument based on good points of unconditional behaviour. Basic principle of proposed egress behaviour: priority to the highest severity of marking. The draft should be discussed on the working group mailing list. ------------------------- o Rethinking TCP-friendly Matt Mathis discussed a presentation he plans to make in the TSVAREA meeting later this week (no slides). The objective of the presentation there is advocate for moving responsibility for fairness into the network. What is the overlap with PCN interests? > Bob Briscoe: one of the research activities is defining what congestion should be. Meeting adjourned at 11:00 am.