Decentralized Resource Allocation in Application Layer Networks

1. Decentralized Resource Allocation in Application Layer Networks T. Eymann, M. Reinicke University Freiburg, Germany O. Ardaiz, P. Artigas, F. Freitag, L. Navarro Polytecnic University Catalunya, Spain

2. Outline • Motivation • Catallaxy Paradigm for Decentralized Resource Allocation • Experiments • Results • Open Issues & Further Research

3. S S S S S Application Layer Network Deployment Application Layer Network (Web Proxy Caching Hyrarchy): 6 servers each requires 1 Mbits net capacity, 200 Mbytes Storage, Less 2 hops from demand regions: A,B,C,D,E S S S S S Resource Allocation D D S S S S S D S D S S D S S S • Programmable Infrastructure: • 30 nodes distributed throught Internet each 10 Mbit net capacity, 2 GByte Storage S S S S

4. Resource Allocation Problem • Centralized RA is computationally intensive (and single point of failure). • And it will get works: • Very Dynamic Infrastructures (Resource nodes come and go frequently): dial up nodes, mobile nodes, ... • High Node Density Infrastructures (Many nodes with little resources): P2P systems, pervarsive computing,..

5. Solution: Economic Markets • Resource Allocation works in Real World with an economic model: allocation of goods among human beings takes place in “markets”. • Markets: • just distribution of utility by a central arbitrator (centralized economy) • decentralized action of utility-maximing agents using a central auctioneer • direct agreement between negotiating agents (Catalaxy)

6. The Catallaxy as a concept for market coordination • Catallaxy is an alternative word for „market economy“ (Mises and Von Hayek of the Neo-austrian economic school) • “Fundamentally, in a system in which the knowledge of the relevant facts is dispersed among many people, prices can act to co-ordinate the separate actions of different people in the same way as subjective values help the individual to co-ordinate the parts of his plan.” (Friedrich A. von Hayek, The Use of Knowledge in Society, 1945) • “The Market” as a technically decentralized, distributed, dynamic coordination mechanism: • Adam Smith’s “invisible hand”, Hayek’s “spontaneous order”, Walras’ “non-tâtonnement process” • Coordination and a stable environment are emergent features of the market • Pursuing local goals alone already stabilizes and coordinates the system.

7. How to Implement Catalaxy: Agents Reasoning, e.g. calculation of a counter-offer using heuristics (may become arbitrarily complex, e.g. AI) Agent Effector, e.g. sent offers (Intention: increase own utility) Sensor, e.g. received offers Environment, e.g. Market

8. Agent-mediated digital economy Characteristics for the agent-mediated digital economy: • Software agents act selfish, because their human owners do: Competition is the norm. • Software agents keep their utility function private: If made public, the agent can be exploited. • Software agents communicate directly: Centralized control institutions can always be bypassed. • Consequences: • Cooperation is always pareto-eliciting (increases utility of all participants) • No free lunch: everyone has a utility function (business model), even centralized institutions • Information is not free or public (every participant operates on private knowledge and subjective values)

9. Negotiation Protocol - Example Client SC Buyer Seller cfp (service access) propose (service access, pS=$24) propose (service access, pB=$18) propose (service access, pS=$21) accept-offer(service access, pB=$21) commit (service access, pS=$21) time time

10. Heuristic-Adaptive Reasoning:Example for a Seller (1) propose (service access, pB=$18) propose (service access, pS=$24) Update Market Price Valuation

11. Heuristic-Adaptive Reasoning:Example for a Seller (2) propose (service access, pB=$18) propose (service access, pS=$24) Should I leave the negotiation?

12. Heuristic-Adaptive Reasoning:Example for a Seller (3) propose (service access, pB=$18) propose (service access, pS=$24) Should I leave the negotiation? Yes reject No Should I make a concession?

13. Heuristic-Adaptive Reasoning:Example for a Seller (4) propose (service access, pB=$18) propose (service access, pS=$24) Should I leave the negotiation? Yes reject No Should I make a concession? No propose (service access, pS=$24) Yes What amount should I concede?

14. Heuristic-Adaptive Reasoning:Example for a Seller (5) propose (service access, pB=$18) propose (service access, pS=$24) Should I leave the negotiation? Yes reject No Should I make a concession? No propose (service access, pS=$24) Yes propose (service access, pS=$21) „costs of life“ (tax) will be deducted in discrete time slots

15. Heuristic-Adaptive Reasoning:Parameters Concession Probability Application Concession Amount Mark-up Continuation Probability Market Price Learning Weight Coordination Negotiation Strategy: Achieving utility maximization setting e.g. concession rate, concession amount, time pressure in relation to market (and the transaction partner). Cooperation Communication Application Services Network Services Physical Services

16. Heuristic-Adaptive Reasoning: adaptation by evolutionary learning  Send „plumage“ (profitx, Genotypex)  select Genotype (profitx)  Create agent (Genotype  Genotype1)

17. Experiments • Simulated Scenarios • Evaluated Dimensions

18. Simulated Application Scenario How to match a network of clients and services? 1 2 3 Acrobat Service Copy of Document MyCompanyPortfolio.pdf (6 Mbytes) Web Server with limited Resource (4 – 60 Mbits) Clients (ADSL 1 Mbit)

19. Catallactic Message Flow Client request_Service (MyComPortfolio.pdf) BW Negotiation Service Negotation

20. Baseline Message Flow Client request_Service (MyComPortfolio.pdf) Master Service Copy as Centralized Auctioner for BW and SC

21. Changing node dynamics high networks It is required an “abstract” simulator , P2P hoc - , ad overloaded Mobile medium CDN networks GRID node density Fixed low medium high CDN P2P Stable A few, powerful A lot, modest GRID Evaluation Dimensions

22. Simulator Scenarios: Resource Density Variations Low Density: Few nodes (5) Lots Resources per Node (60 Mbits) Middle Density: More nodes (25) Less Resources per Node (12 Mbits) High Density: More nodes (75) Less Resources per Node (4 Mbits)

23. Simulator Scenarios:Dynamic Values Very dynamic: Nodes up & down with 40 % probability every 200 ms. Dynamic: Nodes up & down with 20 % probability every 200 ms. Quasi-static: Nodes always up.

24. Simulator - Demand • Clients located in every edge node. • Client request_Service (1 Mbit Server Net Bandwidth, 50 sec). • Random values: • # of demands (among clients) • # of serviceIDs (among 50 diferent videos) • time betwen demands (average 2000 ms) • Moving clients: • Movement time (How often demand moves) • Movement radius (How far demand moves) • Movement percent (How much demand moves)

25. Simulator Choice • The Catnet simulator is build over JavaSim [Univ. Ohio]: JavaSim is a network simulator based in autonomous components. • Javasim implemented in java=>Ease of development, and efficient []. • Javasim models every aspect of a real network: latency, bandwith, lost packets, routing,=>We take into account resource locality (vs. MAS simulators) • Application module implement interfaces of common Inet protocols: TCP, UDP, Mcast =>our components can be modified to work in real world without modification.

26. Preliminary Results • Evaluation Criteria. • Preliminary Results: • Comparison by Scenarios, • Adaptability Evaluation.

27. Evaluation Criteria • RAE (Resource Allocation Efficiency) • The ratio of matched transactions divided by the number of all proposals: # "accepts“/ #"proposals“ • REST (Response Time (Service Access Time)) • How long does it take on average to fill a request:time between “cfp” and “accept” • CC (Communication Costs) • How much communication is needed until the result: # messages * # hops.

28. RAE at quasi-static, slow scenarios Results by criterion – RAE (%) RAE better @ very dynamic Scenario Topology Dependency @ middle density Catallactic Baseline

29. Results by criterion – REST(ms) REST is higher for catalactic: but not as much as expected.GOOD Catallactic Baseline

30. CC increases with density, since higher density means more nodes to send to. Results by criterion: CC (# messages * #hops) CC is similar. But it was expected to be higher because of more negatiations messages: GOOD. Catallactic Baseline

31. Results by Scenario Communicationcost ResourceAllocationEfficiency • Quasi-static • High node density • Very dynamic / low ND • Very dynamic / high ND • Green: confirmed, Red: rejected Reactiontime System b b b b b b c b b b c b

32. Adaptation: Baseline Simulation In baseline system prices keep constant => no adaptation

33. Adaptation: Catallactic Simulation In catalactic system prices adapt over time

34. Open Issues & Further Research • Oscillations, Caotic behaviour. • Tragedy of commons. • Malevolous agents. • Scalability, dynamics. • Theoretical Modelling. Colaboration with agent researchers Colaboration with Complex Adaptive System researchers. • Implementation in grids & P2P scenarios. Colaboration with Grid / P2P projects

35. Thank you, Questions? • More info: • http://research.ac.upc.es/catnet/