19th NMRG-Meeting (January 13/14, 2006, KTH Stockholm) ====================================================== Attendees: ---------- 1. Laurent Andrey (LORIA-INRIA, France) 2. Javier Baliosian (Ericsson, Ireland) 3. Kyrre Begnum (University College Oslo, Norway) 4. Mark Burgess (University College Oslo, Norway) 5. Guillaume Doyen (LORIA-INRIA, France) 6. Alberto Gonzalez Prieto (KTH, Sweden) 7. Lisandro Granville (Federal University Rio Grande do Sul, Brazil) 8. Siri Fagernes (University College Oslo, Norway) 9. Olivier Festor (LORIA-INRIA, France) 10. Markus Fiedler (Blekinge Institute of Technology, Sweden) 11. Johan Nielsen (Ericsson, Sweden) 12. Giorgio Nunzi (NEC Europe, Germany) 13. Aiko Pras (University of Twente, The Netherlands) 14. Matthias Schmid (Infosim, Wurzburg, Germany) 15. Jurgen Schonwalder (International University Bremen, Germany) 16. Rolf Stadler (KTH, Sweden) 17. Radu State (LORIA-INRIA, France) 18. Heimir Sverrisson (Reykjavik University, Iceland) 19. Robert Szabo (BUTE, Hungary) 20. Bert Wijnen (Lucent Technologies) 21. Fetahi Wuhib (KTH, Sweden) Agenda: ------- Thursday (2006-01-12) 09:00 Introduction to Promise Theory Mark Burgess (University College Oslo, Norway) 10:30 Coffee Break 11:00 Promises and Game Theory Mark Burgess (University College Oslo, Norway) 12:30 Lunch Break 13:30 Promises and Prototyping Kyrre Begnum (University College Oslo, Norway) 14:00 Discussion 14:30 Example Pervasive Computing Siri Fagernes (University College Oslo, Norway) 15:00 Coffee Break 15:30 Discussion Everybody 19:00 Dinner (details to be announced) Friday (2006-01-13) 09:00 Traditional Approaches to Distributed Management and their Standardization Jurgen Schonwalder (International University Bremen, Germany) 09:30 New Approaches to Achieve Scalability and Robustness Rolf Stadler (KTH, Sweden) 10:30 Coffee Break 11:00 Examples of Scalable Monitoring through Decentralization Real-time Views using Distributed Query Processing Distributed Threshold Detection Distributed Real-time Monitoring with Accuracy Objectives Alberto Gonzalez (KTH, Sweden) Demos at KTH Networking Laboratory Various speakers from KTH 12:30 Lunch Break 13:30 Invited Presentations Distributed Inter-domain Management: Domain Composition Robert Szabo (BUTE, Hungary) Distributed Configuration and Load-Balancing for Wireless Networks Giorgio Nunzi (NEC Europe, Germany) A Self-organizing P2P-based Framework for Distributed Network Management Markus Fiedler (Blekinge Institute of Technology, Sweden) Matthias Schmid (Infosim, Wurzburg, Germany) 15:00 Coffee Break 15:30 Discussion 16:30 Workshop closes Material: --------- The slides presented during the meeting are all available online at . Minutes: -------- The final editing of the minutes was done by Juergen Schoenwaelder based on notes taken by students from Oslo and Stockholm namely Javier Baliosian, Siri Fagernes, and Kyrre Begnum. First Day: Promise Theory (organized by Mark Burgess) ===================================================== 09:10: Meeting opened by Juergen Schoenwaelder and Rolf Stadler - Welcome and administrative issues (roll call, minute takers). - Overview of the local infrastructure. -------- 09:25: Introduction to Promises (Part I) Presented by Mark Burgess - Mark defines policies as a collection of promises. This approach allows to avoid pitfalls of the policy approach (he believes). - Goal is predictability in the sense of a "flexible" form of stability, perhaps it is predictable stability? Need to accept uncertainties that exist in real-world systems. - Mark questions the Event-Condition-Action (ECA) approach, more precisely the time-scale underlying the policy control loop. - Mark introduces basic ideas behind promise theory (combination of graph theory and set theory). - Q: Don't you end up with something that is a hundred promises long? A: ALL promises may not be practical/possible to describe and or promise because it is not all in our powers. Must accept uncertainty! Where do we draw the line for what we accept as good enough? - Q: Are the links physical? A: The links can be of any kind. Logical or physical. This depends on your modeling goal and resolution. Need modeling tools for complex systems! - Q: What about the propagation of promises? A: Implicit promises: some promises might imply other promises. - Q: A different model. A node has a list of services and a access control list. Are we talking about the same thing? A: The situation might be modeled by promise theory. Different classes of uses can be modeled by different promises. - Q: Is there are way to model conditionals in cooperation, or negative promises? A: It should be doable, since it is possible to imagine in the real work. For negative promises, we might see some example of it, but it should be possible. It is true, that it is hard to verify a NOT-promise. - Q: You did analysis with economies. We in the world follow rules. How do you model rules that are almost "obligations" to society? - A: Part of the next talk. - Q: Stability in decentralized systems is a big problem. Promise theory does not give the answer, can not guarantee stability? A: Yes, we can. Game theory explains how. Nobody can predict stability, but we can predict what kind of promises you need for the probability of stability to be high enough. - Q: I do not really see the advantage of how this idea is better than traditional formal methods. A: This theory is not necessarily different from formal modeling. This is a framework that supports several subjects. This is a starting point of philosophy (autonomy). It is not about the graphs or the notation, but the philosophy. Other (formal) methods tend to go too far in their assumptions. - Q: The notion of promise fits better than agreement. What about broken promises? A: Excellent questions. There are methods for this and it is very interesting. -------- 11:05: Econometric Promise Theory Presented by Mark Burgess - Promises have value, in terms of what is received, what it costs to implement etc. - This involves an exchange of trust, so could a promise be exploited? - Asymmetrical relationships are not stable, as nodes then would either exploit others or be exploited. What is the currency of exchange? - Cooperation - bargaining (basis for understanding the probability of cooperative behavior). - Q: Regarding the adjacency matrix. What about the probabilities? How are they arrived and how reliable are they? A: Two ways: 1. Based on the value of the game (payoff matrix). 2. Simply an observational and derivative approach. But, the identification of an important node and promise might enable us to improve the reliability of the other nodes. There is a real challenge in how to arrive at a matrix of probabilities. - Q: You said this had interesting appliances to security. How? A: Identifying vulnerable spots in the network graph. Look for barriers of communication. Where should you invest the most effort? Identifying the weakest link. - Q: In many cases it might be very difficult to assess the structure of the graph? A: This does not solve all your problems, but you could model certain problems. We must accept that we can not predict everything. - Q: Is this not close to data-mining efforts? A: Based on link analysis, pattern recognition etc. - Q: What is trust? A: Trust is the probability you perceive that the promise will be kept (experience is the rate that promises have been kept in the past). - Q: Promise is a strategic object? A: No, its a local object which may be the same for all the nodes if they pledge allegiance to the same thing. Common goals are an important basis for cooperation. -------- 13:35: Reflection on Promise Theory using Prototyping Presented by Kyrre Begnum - Motivation: How can we investigate / prove certain properties of the theory? How can we identify any hidden assumptions? - Using Maude, a language for executable formal specifications. - Q: Which part of this initiates the interaction/promise? A: This is not covered by promise theory, should be implemented elsewhere. It probably does not matter how the negotiation starts. - Q: How long does it take to evaluate a configuration? A: First it has to be terminating and if you search the whole space it might take quite some time since the search complexity is exponential. - Q: Do you not run into the same problems as other formal approaches? A: It seems that proving Needham-Schroeder is already something difficult to do without specific optimizations. - Q: Would narrowing help? A: Mark says the purpose of the exercise is to verify some simple examples and then go back to heuristics to address more complex promise graphs. -------- 14:15: A Smart Mall Scenario Using Promise Theory Presented by Siri Fagernes - Outcome is very sensitive to the chosen metrics and values - Analysis shows that not only policy that you decide yourself is important but also the emerging behavior of the system. - Q: The adjacency matrix of probabilities has values > 1. Why? A: This one is a sum of all probabilities to all promises. Second Day: Distributed Management (organized by Rolf Stadler) ============================================================== 09:00: Traditional Approaches to Distributed Management and their Standardization Presentation by Juergen Schoenwaelder - Juergen present a classification of management paradigms and presented some approaches. - On Active Networks Juergen stated that they are of academic interest but not practical, Rolf disagreed in the case of bounded and well defined applications. - On Remote Operations MIBs, Juergen added that they are being extended and that they have proven useful for operators. - Rolf asked if Expression MIBs are management by delegation. Answer: no, there is a specific section in the slides for this. - On Smart Agents, Juergen states that it is a great standardization failure because nobody is using it. - Discussion on point 1 of Next Steps in Distributed Management: Olivier says that GRID people believe that there are no resources available for management in the nodes. Rolf says that Cisco says that there would be resources but constrained. Aiko: there is no willingness to give resources for management, but if you have a good use case would be. Juergen: Is not a technical problem, it is economical. Bert: the people that have the knowledge don't have the time. - Markus Fielder asked about which approach is RMON MIB. RMON is Approach #0 because local instrumentation of nodes. Rolf stated that RMON is Distributed Monitoring but restricted to a given media not to a whole distributed system. -------- 09:30: New Approaches to Achieve Scalability and Robustness Presentation by Rolf Stadler - Reference papers at http://www.ee.kth.se/~stadler/nmrg/ - Rolf speaks about today's networks, they are not one service oriented and are becoming less robust and scalable in terms of management. He speaks about distributed management, the reasons of its failure until now and a new p2p-like approach by KTH (Weaver devices). Mark says that the idea is close to cfengine and speaks about behavior learning on the higher layers. - Rolf speaks about DHT for routing of management messages but wonders if an epidemic approach may be cheaper. - Mark have a paper on sampling techniques. Telenor is interested on epidemic approaches. - Juergen asks if you can do arbitrary function processing with the echo algorithm or whether you can only do things like max min average etc. Rolf replies that you CAN do arbitrary queries. -------- 11:00: Weaver Query System and Weaver Query Language Demos by Alberto Gonzalez Prieto - Alberto shows a demo on Weaver Query System and Weaver Query Language: There is no knowledge on how this would work on a gigabit network with heavy traffic. - Bert Wijnen points that the serial CLI interface used would be too slow in this case. -------- 11:30: Distributed Real-time Monitoring with Accuracy Objectives Presentation by Alberto Gonzalez Prieto - Alberto presents A-GAP, a distributed, asynchronous protocol for monitoring (reference paper at http://www.ee.kth.se/~stadler/nmrg/AGAP-KTHTR-2005.pdf). - Aiko will provide traces of real traffic to give real usage case. The error estimation changes with the dynamics of the network (come and go of nodes and reorganizations). - Rolf did not find literature on accuracy of aggregate monitoring. Markus has some work on that, he will send it to Rolf. -------- 13:30: Decentralized Computation of Threshold Crossing Alerts Presentation by by Fetahi Wuhib - Fetahi presents a decentralized algorithm to compute threshold crossing alerts. The talk is based on a paper published at DSOM 2005: http://www.ee.kth.se/~stadler/nmrg/TCAGAP-DSOM2005.pdf - Fetahi presents TCA-GAP a GAP-based protocol for scalable monitoring of threshold crossing alerts on aggregated parameters. - The presentation uses SUM as an example and Javier asks about other aggregations. Rolf answer that there exists an algebra to specify how distributed thresholds are computed for any aggregation. - Rolf adds that there is theoretical ways to compute how long a spike in the measured parameter should be for allowing the system to detect this variation. - Fetahi made a demo of the TCA-GAP protocol using KTH's testbed. -------- 14:00: Distributed Inter-domain Management: Domain Composition Presentation by Robert Szabo - Robert presents shortly the Ambient Networks project and its P2P approach and the work on domain composition they are carrying out. - They are the only ones working on domain composition in the Ambient Networks project. -------- 14:30: Distributed Configuration and Load-Balancing for Wireless Networks Presentation by Giorgio Nunzi - Giorgio presents work from NEC on self-configuring access points, load balancing and monitoring of self-configuring devices. - Synchronization between APs is a very important issue. - Giorgio shows convergence time for bootstrap and spread time of new global information based on simulations. - They use versioning for information distribution to avoid loops. - Discussion about why not solve the problem by over-provisioning due the low price of APs. -------- 15:00: A Self-organizing P2P-based Framework for Distributed Network Management Presentation by Markus Fiedler - Markus presents AutoMon, a system for self-organized QoS monitoring. - P2P network for connecting DNAs (Distributed Network Agents). - They use Kademlia as location protocol. - Matthias Schmid made a demo on AutoMon.