PhD Thesis Defence: Managing Incentives in Community Network Clouds

PhD Thesis Defence: Managing Incentives in Community Network Clouds
Date: 21 June 2016
Venue: Department of Computer Architecture, Universitat Politècnica de Catalunya. BarcelonaTech. Barcelona, Spain.

Abstract:
Internet and communication technologies have lowered the costs for communities to collaborate, leading to new services like user-generated content and social computing, and through collaboration, collectively built infrastructures like community networks have also emerged. While community networks focus solely on sharing of network bandwidth, community network clouds extend this sharing to provide for applications of local interest deployed within community networks through collaborative efforts to provision cloud infrastructures. Community network clouds complement the traditional large-scale public cloud providers similar to the model of decentralised edge clouds by bringing both content and computation closer to the users at the edges of the network. Community network clouds are based on the principle of reciprocal sharing and most of their users are moved by altruistic principles. However, as any other human organisation, these networks are not immune to overuse, free-riding, or under-provisioning, specially in scenarios where users may have motivations to compete for scarce resources. We focus in this thesis on the incentives based resource regulations mechanisms to derive practical ways of implementing arbitration when such contention for limited resources occurs. We first design these regulation mechanisms for the local level where stronger social relationships between the community members imply trust, and ensure adherence to the system policies. We next extend the mechanisms for larger communities of untrusted users, where rational users may be motivated to deviate for their personal gains, and develop a distributed framework for guaranteeing trust in the resource regulation. Such mechanisms assist in encouraging contribution by the community members, and will help towards adoption, sustainability, and growth of the community cloud model.

Transcript

1. Managing Incentives in Community Network Clouds Amin Khan amin.khan@ieee.org / aminmkhan.com Advisors: Felix Freitag (UPC), Luís Rodrigues (INESC-ID, IST) Erasmus Mundus Joint Doctorate in Distributed Computing (EMJD-DC) 21 June 2016 Barcelona
2. Outline • Context • Incentives • Incentive Compatibility • Trust • Conclusion 2
3. Community Networks • Social collective to build bottom-up communication infrastructures • Scalable, self-organized and decentralized IP networks operated by citizens 3
4. Community Network Clouds • Cloud infrastructure and services for community networks – built within community networks – hosted on community-owned computing and communication resources – providing services of local interest 4
5. Objective • To incentivize contribution and maximise the utility of the community cloud system, while guaranteeing trust and ensuring truthfulness from its participants 5
6. Methodology • Simulation experiments • Prototype deployments in Clommunity testbed 6
7. Contributions • For local communities of trusted users – Incentive-based resource regulation mechanisms • For community of untrusted users – Novel distributed virtual auctioneer framework 7
8. Results Resource Regulation Distributed Auctioneer Community Network Cloud Hardware and Networking Layer Core Layer Services Layer Front End Layer Ch. 5 § 6.1.3 Ch. 4 Trusted Users Untrusted Users § 6.1.1 Ch. 3 DSS § 6.1.4 Middleware Layer § 6.1.2 § 6.1.5 8 Framework Incentives Trust Ramifications and Collaborations
9. Community Cloud 9
10. Federated Community Clouds Local Community Cloud Super Node (SN) Ordinary Node (ON) 10
11. Cloudy — OS for Cloud in a Box 11 R Baig et al. The Cloudy Distribution in Community Network Clouds in Guifi.net. IFIP/IEEE IM, 2015. http://cloudy.community
12. Architecture for Community Cloud 12 Incentives Trust CLI GUI API Front End Layer Core Layer Hardware Layer Middleware Layer Services Layer Storage Video Network Cloud Services Cloud Coordinator Resource Regulation Trusted Auctioneer Support Services VM Controller VM Monitor VM Scheduler Management Services VMs VMs VMs
13. Challenges in Community Clouds • Independent Providers • Lack of Centralized Control • Incentivizing Contribution • Community of Untrusted Users • Network and Platform Heterogeneity • Latency and Contention in Mesh Network
14. Outcomes • Community Clouds – AM Khan, F Freitag, L Navarro – Encyclopedia of Cloud Computing, 2016 • Current Trends and Future Directions in Community Edge Clouds – AM Khan, F Freitag, L Rodrigues – IEEE CloudNet, 2015 14
15. Outcomes – Architecture • Cloud services in the Guifi.net community network – M Selimi, AM Khan, E Dimogerontakis, F Freitag, and R Pueyo – Computer Networks, 2015. [JCR IF: 1.256, Q2] • Towards Distributed Architecture for Collaborative Cloud Services in Community Networks – AM Khan, M Selimi, F Freitag – 6th International Conference on Intelligent Networking and Collaborative Systems (INCoS), 2014 15
16. INCENTIVES 16
17. Incentives • CNs follow reciprocal sharing principle – Wireless Commons License – Pico-Peering Agreement • Supported by social context of the community • Simple social, hedonic & psychological motivations • Existing incentives sufficient for Community Clouds? 17 M. Bina, and GM Giaglis. Unwired Collective Action: Motivations of Wireless Community Participants. IEEE ICMB, 2006.
18. Reciprocity Based Incentives Contribution-Based You get what you pay! Biased toward nodes with higher capacity Fairness? Effort-Based Parecon Principle Give everyone equal chance to participate No matter what their capacity In effort-based, weigh in users' capacity in addition to their contribution Capacity Variability Heterogeneity R. Rahman et al. Improving Efficiency and Fairness in P2P Systems with Effort-Based Incentives, IEEE ICC 2010.
19. Incentives-Based Resource Regulation • Nodes request resources • Check node’s credit • Resources available in local zone • If not, request resources from other zones • Allocate resources • Assign/Deduct credits • Credits ≈ Contribution / Capacity 19
20. Experimental Setup • Simulation experiments with Python • Federated Community Cloud with 100 zones • Nodes with different capacity, sharing behavior Total Capacity Shared Capacity
21. Sharing Behaviour & Success Ratio 21 Sharing benefits irrespective of the capacity constraints on contributions.
22. Requests Fulfillment 22
23. Discussion • Regulation helps motivate contribution • Computationally efficient, minimal overhead • Reliant on users reporting their true capacity • Users can hide capacity to gain advantage • Requires a central and trusted allocator • Careful calibration to avoid discouraging high contributors 23 but
24. Outcomes • Prototyping Incentive-based Resource Assignment for Clouds in Community Networks – AM Khan, U Buyuksahin, F Freitag – 28th IEEE International Conference on Advanced Information Networking and Applications (AINA ‘14) – Best Paper Award [Core Rank B] • Towards Incentive-based Resource Assignment and Regulation in Clouds for Community Networks – AM Khan, U Buyuksahin, F Freitag – 10th International Conference on Economics of Grids, Clouds, Systems, and Services (GECON ‘13) 24
25. Outcomes • Incentive-based resource assignment and regulation for collaborative cloud services in community networks – AM Khan, U Buyuksahin, F Freitag – Journal of Computer and System Sciences, 2014. [JCR IF: 1.138, Q2] 25
26. INCENTIVE COMPATIBILITY 26
27. Challenges in Resource Allocation Maximal Social Welfare Ex-Post Budget Balance Truthfulness Computational Efficiency Trust in Allocator
28. Bandwidth Reservation at Gateways Optimal Truthful
29. Static Pricing Mechanisms • Fixed Usage-Based Pricing – Everyone pays the same, including for free • Priority Pricing – Pay according to the priority of the requests • Figure out price values with varying demand? • Prevent everybody asking for high priority? Challenges P Maillé and B Tuffin. Telecommunication Network Economics: From Theory to Applications. 2014
30. Dynamic Pricing Mechanisms • First Bid (Sealed-Price) Auction – Everyone quotes a price reflecting its valuation – Winning users pay the amount they bid • Generalized Second Price (GSP) Auction – Winner pays the next highest bid price – Often used to sell keyword-based Internet ads • Vickrey–Clarke–Groves (VCG) Auction – Ensures truthfulness and maximal social welfare P Maillé and B Tuffin. Telecommunication Network Economics: From Theory to Applications. 2014
31. Priority Classes Low High Differentiated using step function, based on request waiting time.
32. Loss in Social Welfare Lying Hurts! Truthfulness Critical for Sustainability
33. Discussion • Without truthful responses from participants, utility of the system suffers • Gratis, and Static pricing based mechanism – Not suitable, can’t ensure truthfulness • Second-price auctions serve well! – VCG for long-term allocation blocks – GSP for regular allocations, as is faster • But require “Trusted Auctioneer”
34. Outcome • Towards Incentive-Compatible Pricing for Bandwidth Reservation in Community Network Clouds – AM Khan, X Vilaça, L Rodrigues, F Freitag – 12th International Conference on Economics of Grids, Clouds, Systems, and Services (GECON ‘15) 34
35. TRUST WITH DISTRIBUTED AUCTIONEER 35
36. Community of Untrusted Users • Most literature on auction mechanisms for optimal and truthful resource allocation assumes a centralized trusted auctioneer • CNs lack such a central entity, which even if it existed, can affect scalability • With multiple rational service providers in CNs, users can’t trust them as auctioneer(s) • Users can be – Altruistic or Acquiescent – Byzantine – Faulty or Malicious – Rational 36
37. Issues with Trust in Auctioneer • Consider three providers – Alice, Bob, Mallory • Mallory perturbs the bidding process to its advantage – Can select winning bids only for itself – Select higher or lower price to its advantage – Mallory can make coalition with Bob to defeat Alice • Synchronous and Asynchronous scenarios • Coalitions of rational players • Solution involves replication and cross-validation 37
38. Distributed Auctioneer • Novel framework for devising distributed simulations of auctioneer, which are • Nash equilibria k-resilient (ex post) to asynchrony and coalitions of rational providers of size at most k, and • Leverage the distributed nature of the virtual trusted entity to parallelize the resource provisioning algorithms 38
39. Components • Input Validation • Bid Agreement • Common Coin • Data Transfer 39 Abraham et al. Distributed Protocols for Leader Election: A Game- Theoretic Perspective, DISC 2013. Afek et al. Distributed computing building blocks for rational agents, PODC 2014.
40. Resource Allocation Instances • Standard Auction – Based on VCG mechanism • Double Auction – Based on McAfee mechanism 40 X. Zhang et al. A Truthful (1-ϵ)-Optimal Mechanism for On-demand Cloud Resource Provisioning, INFOCOM 2015. Zheng et al. STAR: Strategy-Proof Double Auctions for Multi-Cloud, Multi- Tenant Bandwidth Reservation, IEEE Trans. on Computers, 2014.
41. Distributed Auctioneer Framework 1 2 m B A A 1 2 m bidders submit bids bidders collect results Providers Providers . . . . . . Framework bm b1 b2 b b b (x,p) (x,p) (x,p) 41 Bid Agreement Allocation
42. Task Decomposition T3 T1 T2.1 T2.2 42 Execute in Parallel
43. Parallel Allocator Framework 1 2 4 Providers Allocator b b b (x,p) (x,p) (x,p) T1 3 1 2 4 3 (x,p) T1 T2.1 T2.1 T2.2 b T2.2 D T T3 T3 T3 T3 CC I V T1 T1 Providers 43 Common Coin Input Validation Data Transfer
44. Experiment Setup • Prototype implemented in Python – PyPy compiler – ZeroMQ messaging library • Clommunity Testbed in Guifi.net CN • 4 Nodes in Catalonia, Spain – 2 in UPC, 1 in Hangar, 1 in Taradell – Intel Core i7, 16 GB RAM, 1 TB machines – Debian 7 x86 1 CPU, 2 GB RAM, 10 GB OpenVZ containers 44
45. Double Auction 45 Minimal overhead when using Distributed Auctioneer
46. Standard Auction 46 Parallelization provides speed-up despite the communication overhead
47. Discussion • Overhead of the distributed auctioneer is not significant • Supports parallelization for substantial gains in performance • Applicable to rational providers only • Not resistant to faulty or malicious providers 47 but
48. Outcome • A Distributed Auctioneer for Resource Allocation in Decentralized Systems – AM Khan, X Vilaça, L Rodrigues, F Freitag – 36th IEEE International Conference on Distributed Computing Systems (ICDCS 2016) [CORE Rank A] 48
49. Conclusion • Incentive-based Regulation Mechanisms – Adapted to social, economic and technical context • Distributed Auctioneer Framework – Efficient and Optimal allocation with Trust • Building blocks for resource allocation in community cloud system 49
50. Open Issues & Future Work • Fail-Stop Model • Byzantine Users • Network Topologies • Multi-Resource Allocations 50
51. Thank You! Amin Khan @aminmkhan