Congestion & Pre-congestion Notification (PCN) Working Group Meeting Minutes IETF 69 - Chicago July 25, 2007 9:00am-11:35am CDT Meeting minutes taken by Bruce Davie Chairs - Scott Bradner & Steven Blake o Administrivia The agenda was presented and the order of two presentations (dratf-chan-pcn-encoding-comparison-00 and dratt-westberg-pcn-load-control-00) were switched. The working group milestones from the charter were reviewed. Steve Blake suggested some questions participants should ask about assumptions made in the various approaches (see slide). o Phil Eardley: Pre-Congestion Notification Architecture draft-eardley-pcn-architecture-00.txt Aiming to make this a WG draft to meet the 1st milestone (info RFC) includes broad set of authors Draft aims to be compatible with many approaches to marking and edge behaviors> Numerous comments have been received requesting clarifications on terminology, scope, design goals, deployment scenarios. Need to add: - single marking is OK - tunneling discussion - addressing - probing & ECMP OAM needs review - Tom Taylor will comment by end of week Tina Tsou will send some OAM comments to the list David Black points to RFC 2983 - Diffserv & Tunnels Bob Briscoe mentions ECN tunneling in 3168 and a new draft (draft-briscoe-tsvwg-ecn-tunnel-00.txt) on ECN tunneling - he will help with text for PCN arch. Anna Charny on terminology 4th option - naming could be based on semantics rather than function. Kwok Ho-Chan Supports Anna's view. Anna Charny suggests - "upper" and "lower" threshold. Joe Babiarz - suggests we move on. Georgios Karagiannis - supports Anna. Bruce Davie asked for show of hands to see if there is broad support for upper/lower - yes, there is. Tina Tsou - Does ingress need to know address of centralized node? Steve Blake says there is no centralized node. Anna Charny - current draft does include concept of central node. Tom Taylor - central node has benefits; "quicker" solution. Joe Babiarz- should focus on determining ingress and egress nodes for a flow Bob Briscoe - trying to address how PCN can fit into existing systems that have a central PDP. Kwok Ho-Chan - central PDP a deployment scenario for future consideration. Lars Eggert (TSV AD) - charter is clear in that egress should be able to make decision. Bob Briscoe - some proposals have ingress making decision. Georgios Karagiannis - PCN should not focus on finding address of central node. Lars Eggert - would like to avoid mismatch between devices that must be centralized and devices that must work at egress. Steve Blake - ECMP handling has implications on flow state. Bob Briscoe - when a node fails, we know it will cause problems in ECMP - would be effectively like no PCN. Anna Charny - worst case is bad - should look at some real data. Joe Babiarz - there is a valid proposal for ECMP treatment. Scott Bradner - show of hands - working group document? strong support by the meeting participants. There are too many authors, if all are contributing then select editor and put rest of names in acknowledgements section else if only a few are contributing have them as authors but do not use list of authors to indicate support. o Kwok Ho-Chan: Pre-Congestion Notification Encoding Comparison draft-chan-pcn-encoding-comparison-00.txt Survey, establish criteria for comparison, assist selection. Bob Briscoe - question about number of states - seems to be one too many. Issues - What's the relation between Functional Features and States? Plan to update draft to replace "feature" with "state". Bob Briscoe - Admission marking is a toggling between 2 states. Steve Blake - What is the intent of ADs for this document? Lars Eggert - Goal: explain how to shoe horn the needed signals into the few bits, discuss interactions with other uses (e.g. DSCP). Maybe this draft is too early? Kwok Ho-Chan - Algorithm drafts are trying to be encoding independent. Lars Eggert - Once you know what needs to be conveyed, then figure out how to encode it. Bob Briscoe - Some schemes use more code points than others, and there may be tradeoffs. Lars Eggert - agrees, maybe premature to look at encodings in this level of detail. Georgios Karagiannis - Should we delay this draft? Lars Eggert - Feel free to work on it if you have time to burn. But how useful is it? List of candidate algorithms might be shorter after this meeting. Anna Charny - Some complexity might be avoided. We know how many states in each algorithm. Maybe should focus on how to encode the right number of states. Kwok Ho-Chan - The evaluation criteria are the important part of the draft - please give input. Francois le Faucheur - One approach is to identify the number of code points for a given approach and then discuss how to encode them - will you do this in -01? Kwok Ho-Chan - -01 will be much shorter, Still plans to say what the bit patterns mean. Lars Eggert - If it doesn't matter what info is in what bit, then can just talk about the number of bits, not what they mean. Bob Briscoe - the one reason why it matters is tunnels. Phil Eardley - suggest structuring draft as: If you want n states, here are your options. Anna Charny - For all proposals, it doesn't matter what bits encode what states. How to encode the number of states is what matters. Steve Blake - Asks authors for another revision before WG adoption. Kwok Ho-Chan - Can we make that decision on the mailing list? Still want to meet the WG milestone. Steve Blake - All WG decisions are made on the mailing list. Will send out a call for comments. o Georgios Karagiannis: LC-PCN - The Load Control PCN Solution draft-westberg-pcn-load-control-00.txt Phil Eardley - Are you using the same bit marking to indicate both congestion and severe congestion? Georgios Karagiannis - Yes. Phil Eardley - So, does the rate at which you mark packets change? Georgios Karagiannis - No. Phil Eardley - Marking rate seems to drop as congestion gets worse. Georgios Karagiannis - There are other things going on beside what is shown in the presentation (!!) Phil Eardley - Must Admission control use probing? Georgios Karagiannis - Yes. Anna Charny - There are some things that don't seem right, can they be fixed in the next version? Georgios Karagiannis - Yes, but maybe not before September. o Anna Charny: Pre-Congestion Notification Using Single Marking draft-charny-pcn-single-marking-02 Performance Evaluation of CL-PHB Admission and Termination Algorithms draft-zhang-pcn-performance-evaluation-02 High level comparison of Virtual Queue (VQ) and Single Marking (SM) admission. SM is sensitive to bursts at low aggregation, and over-admits More brittle than VQ when number of flows is in the 1-2 range. Both schemes unfair to long-haul aggregates in the multi-bottleneck case. Bob Briscoe - Unfairness to long flows a feature, not a problem? Anna Charny - It's really quite bad. Steve Blake - Is the effect at low aggregation levels really a problem? Assume large aggregation. Anna Charny - There may be cases where it is a problem, because there are many flows at bottleneck but some I-E flows are not highly aggregated. Bob Briscoe - And that might lead to picking one scheme vs. another. Beat down of long flows with multiple bottlenecks - needs large demand for long period - probably of limited practical impact. David Black - This looks like congestion collapse - goal should be get out of this state. Anna Charny - It's hard to do in this case. There is a constant demand that considerably exceeds the capacity Summary: Both schemes work well in most cases, the salient difference being when small number of flows make up some ingress-egress aggregates. Francois le Faucheur - The SM scheme uses a metering scheme that already exists in routers. o Joe Babiarz: Three State PCN Marking draft-babiarz-pcn-3sm-00 Tom Taylor - Slowdown parameter - is it the same for all nodes? Joe Babiarz - Does not need to be constant - variation affects termination rate; it's a configured parameter. o Jozef Babiarz: Simulation Results for 3sM draft-babiarz-pcn-explicit-marking-01 Tina Tsou - What is the service class BW limit? Joe Babiarz - It's the amount of BW available in the router to forward that class without loss. Bob Briscoe - See his other draft in tsvwg on small/big packets draft-briscoe-tsvwg-byte-pkt-mark-00.txt.