Energy-aware Service Allocation for Cloud Computing Tobias Widmer 1 - PowerPoint PPT Presentation

Energy-aware Service Allocation for Cloud Computing Tobias Widmer 1 , Marc Premm 1 , Paul Karänke 1,2 1 Universität Hohenheim 2 FZI Forschungszentrum Informatik, Karlsruhe University of Hohenheim Information Systems 2

Outline 1. Research Approach 2. Related Work 3. Model 4. Allocation Mechanism 5. Evaluation 6. Conclusion T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 2

1. Research Approach  Object of research service providers customers Service network of customers and  sp1 cloud service providers (SPs), i.e., c1 multiple SPs offer (standardised) sp2 cloud services to multiple customers c2  Problem sp3 Energy-aware optimal service allocation   Perspective Multiagent system technology   Contribution: Auction-based allocation mechanism as a heuristic (method artefact) T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 3

Basic Assumptions SPs offer (standardised, i.e., equal functional properties) cloud services of  specified energy consumption rate (metric: performance-per-watt) [Sharma et al. 2006] SPs know their energy efficiency and capacity  Energy-efficient cloud services are scarce  Customers outsource infrastructure, platforms, and software applications  to cloud SPs such that physical machines can be switched off (  leading to higher energy efficiency) Customers and SPs model preferences via utility functions  Service cost and energy efficiency correlate positively  service provider customers c1 sp1 c2 T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 4

2. Related Work  Energy-efficient cloud computing  Most approaches consider energy efficiency within a single data centre but do not take a global (social welfare) perspective [Berl et al. 2009; Baliga et al. 2010; Bodenstein et al. 2011]  Global cloud computing architecture on energy efficiency and its impact on carbon dioxide emissions but requires central coordination entity [Garg et al. 2011]  Agent-based cloud service allocation  Central coordination entity required [Parkes & Shneidman 2004]  Distributed coordination approaches do not consider energy efficiency in the allocation rationale [Bo & Lesser, 2010; Bo et al. 2010] T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 5

3. Model E j Energy efficiency of SP in performance/watt W j Capacity of SP in performance c j (s k ) Cost of SP to provide service of type s k V i (h l i ) Valuation of customer for host h l i w(s k ) Computing requirement for service of type s k in performance T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 6

Maximisation Problem  Find allocation x ijk that maximises the sum over all utilities, i.e., utilitarian social welfare [Chevaleyre et al. 2006] W* = max ∑ ij (∑ l (V i (h l i ) x ijk - ∑ k c j (w(s k ), e k )) x ijk ) subject to for all j : ∑ j ∑ k w(s k ) x ijk ≤ W j  Computation problem is NP-complete [Garey & Johnson 1979]  Heuristic is needed to approximate optimal solution  We propose a distributed allocation mechanism T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 7

4. Allocation Mechanism  Combinatorial multi-attribute procurement auction [Parkes & Kalagnanam 2005; Che 1993]  Considers bundles of cloud services  Considers price and energy efficiency as attributes  Request bundle R i = ((r 1 , ..., r n ), (r 1 , ..., r n ), ... , (r 1 , ..., r n )) with r k = 1 iff service s k belongs to requested bundle, 0 otherw.  The mechanism defines a score considering energy efficiency and price: σ i (s k , e k , p k ) = α i (w(s k ) / e k ) – β i p k with scalars α i , β i  Bid bundle B j = (( σ i (s 1 , e 1 , p 1 ), e 1 ), ( σ i (s 2 , e 2 , p 2 ), e 2 ), ... , ( σ i (s n , e n , p n ), e n )). If j does not bid for s k : ( σ i (s 1 , e 1 , p 1 ), e 1 ) := (0, 0) T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 8

Second-score Auction  Second-score auction: winning SP must match second highest score [Che 1993]  Example for a service of type s:  A bids (5, 0.2) with score σ i (s, 5, 0.2) = 10  B bids (6, 0.15) with σ i (s, 6, 0.15) = 9 A is winner and is free to choose any tuple (e’, p’) such  that σ i (s, e’, p’) = 9 T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 9

Auction Protocol service provider customer Announce request for service types Submit binding bids for service types Inform winner and transfer second score Calculate optimal tuple (energy, price) based on second score: final contract is established The proposed protocol is incentive-compatible for SP agents. T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 10

5. Evaluation  Objective: To demonstrate the efficiency of the artefact (nearness to optimum)  Evaluation method: Experimental evaluation – simulation with artificial data (inspired by real-world data)  Experimental setup  From Amazon and SPECpower: 1 Mio. ssj_ops cost 6.75 € with power consumption rate 186 Watt  Poisson dist.: 10 hosts with 2 services/host  10 services: performance ranging from 0.3 to 2.1 Mio. ssj_ops  2..100 customer agents, 5 SP agents, 100 experiments  Supply/demand balance at approx. 50 customer agents  Different weights on energy and price (see score definition) T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 11

Results (1): Energy-aware customers Excess supply: customers Services are scarce: high easily find energy-efficient competition; customers are willing to services but SP costs are high forgo their energy-price preferences T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 12

Results (2): Price-aware customers Low competition: cheapest Services are scarce: high SPs are selected; results in competition; customers are willing to higher social welfare forgo their energy-price preferences T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 13

Conclusion  Findings  A formal, game-theoretic framework for modeling energy-aware cloud service allocation  A novel distributed allocation mechanism that integrates energy efficiency into the allocation rationale  Limitations  Considers two non-functional service properties only (energy efficiency and price)  Only 5 SP agents are considered in the experiment T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 14

Thank you for your attention! Tobias Widmer Information Systems 2 University of Hohenheim tobias.widmer@uni-hohenheim.de Förderkennzeichen 01ME11052 www.migrate-it2green.de T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 15

References Baliga, B.J., Ayre, R.W.A., Hinton, K., Tucker, R.S. (2010). Green cloud computing: Balancing energy in  processing, storage and transport. Proceedings of the IEEE 99, 149 – 167. Berl, A., Gelenbe, E., Di Girolamo, M., Giuliani, G., De Meer, H., Dang, M.Q., Pentikousis, K. (2009). Energy-  efficient cloud computing. Comput. J. 53, 1045 – 1051. Bo, A., Lesser, V. (2010). Characterizing contract-based multi-agent resource allocation in networks. IEEE Sys.  Man. Cybern. 40, 575 – 586. Bo, A., Lesser, V., Irwin, D., Zink, M. (2010). Automated negotiation with decommitment for dynamic resource  allocation in cloud computing. In: Ninth International Joint Conference on Autonomous Agents and Multi-Agent Systems (AAMAS’10), pp. 981– 988. Bodenstein, C., Schryen, G., Neumann, D. (2012). Energy-aware workload management models for operation  cost reduction in data centers. European Journal of Operations Research 222, 157-167. Che, Y.K. (1993). Design competition through multidimensional auctions. RAND Journal of Economics, 24 (4),  668-680. Chevaleyre, Y., Dunne, P.E., Endriss, U., Lang, J., Lemaitre, M., Maudet, N., Padget, J., Phelps, S., Rodriguez-  Aguilar, J.A., Sousa, P. (2006). Issues in multiagent resource allocation. Informatica 30, 3 – 31 . Fischer, K., Müller, J.P., Heimig, I. and Scheer, A. (1996). Intelligent agents in virtual enterprises. In  Proceedings 1st International Conference on the Practical Application of Intelligent Agents and Multi Agents Technology, London. Fuehrer, E. C. and Ashkanasy, N. M. (1998). The Virtual organization: defining a Weberian ideal type from the  inter-organizational perspective. Paper presented at the Annual Meeting of the Academy of Management, San Diego, CA. T. Widmer et al.: Energy-aware Service Allocation for Cloud Computing 16