Applying to Decision- Making in Autonomic Systems Presented by Andres J. Ramirez Authors: Andres J. Ramirez, David B. Knoester, Betty H.C. Cheng, and Philip K. McKinley 1 Motivating Application Remote data mirroring - prevent data loss & data unavailability. - multiple competing objectives [Keeton06, Wilkes03, Wilkes04]: Reliability Performance Cost 2
Motivating Application Consider the following network of remote data mirrors: - what if a link fails? - parts of the network become disconnected - severe financial penalties - network must be reconfigured! 3 Motivating Application How should we reconfigure the network? - must consider tradeoffs... - more than a trillion possible candidates! 4
Main Challenge How can we automatically identify target reconfigurations? - searching the entire solution space may not be practical - must consider competing objectives must deal with uncertainty!environment may change at run time 5 Requirements Automatic generation of new reconfigurations - respond to current environmental conditions - considers tradeoffs between functional and non-functional requirements - new solutions produced and applied in real-time 6
But How? Darwinian evolution has produced countless examples of self-adaptive behavior. - Survival of the fittest guides evolution. We can harness the power of natural selection to evolve target reconfigurations during execution.! leverage genetic algorithms! 7 Traditional Approaches Related Work - balance multiple objectives [Kephart03,Walsh04]. Evolutionary Approaches - genetic algorithms for dynamic networks [Fabregat05, Lu07, Tseng06,Wang08].!methods not applied at run time. - reconfiguration of a network s topology [Montana02].!expensive forms of evaluating solutions.!results did not support online reconfiguration. 8
Evolutionary decision-making approach to evaluate tradeoffs automatically in real-time. Monitor Monitor... Monitor Evolutionary Framework Reconfiguration Monitor Evolve an overlay network to diffuse data across a set of remote mirrors. 9 Start Encode Crossover Individual maps 010111010101 Solution Mutation Evaluation 3.1 4.2 0 5.1 1.2 2.3 Population Selection Best 50% Terminate End [Yes] [No] Mutations New Individuals Individuals 10
Use graph to represent network. -edges represent active/ inactive connections -propagation methods [Keeton04,Keeton06,Wilkes03] Encoding!synchronous: each write must be acknowledged b at destination.!asynchronous: batches of data are sent at periodic intervals. With 25 remote mirrors, d a 7 * 2^300 candidate solutions! Encoding = <ab, = <1, bc, 0, cd, 0, 1, ad, 1, ac, 0> bd> c 11 Crossover Objective: exchange parts of the overall solution -transfer edge properties A A B B D D C C A A B B D D C C Network I Network J Network IJ Network JI Effect: combine key parts of the overall solution to generate fitter solutions. 12
Mutation duce variation in the solution space. -randomly activate / deactivate links -randomly reassign a propagation method A B Network IJ' Effect: explore additional areas of the solution space D C 13 Fitness Functions FF =!1*Fc +!2*(Fe1 + Fe2) +!3*(Fr1 + Fr2) Cost Performance Reliability Fc = 100 - (100 * cost/budget) latencyavg Fe1 =50 - (50 * ) latencywc Fe2 =50 * ( bandsys - bandeff bandsys linksused Fr1 =50 * ( ) linksmax datalosspot Fr2 =50 * ( ) datalosswc + bound) 14
Multidimensional Reconfiguration Original design optimized only for: -cost 12 13 9 8 23 0 2 5 18 11!1= 1,!2 = 0,!3 = 0 - what if a link fails? 4 20 10 6 15 22 19 1 16 17 24 7 21 3 14 15 Multidimensional Reconfiguration 17 1 Reconfigured design optimizes for: -reliability, performance, cost 14 21 15 12 9 3 16 9 13 8 84 23 0 6 24 2 5 15 18 12 13 18 11!1= 1,!2 = 2,!3 = 2 - what if a link fails? 420 10 20 22 22 19 1 16 17 5 19 2 24 7 7 23 21 3 14 10 11 6 0 16
Fitness throughout reconfiguration process -initial overlay fails at generation 2500 /+0123-,)(44 Max. Fitness %"" #"" "!#""!%"" 2 -suitable reconfigurations found by generation 3500 < 30 seconds in a laptop!!""" #""" $""" %""" &""" '()(*+,-.) 3-,)(44 2 17 Number of active links in overlay network -initial overlay design fails with single link Number of Links Number of Links 120 100 failure -reconfigured design increases the number of links in the overlay network. 80 60 40 20 1000 2000 3000 4000 5000 Generation Num. of Links 18
Potential data loss throughout overlay network -initial overlay design did not consider reliability Potential Avg. Data Loss 1.5 1 0.5 -reconfigured overlay design reduces potential data loss 0 1000 2000 3000 4000 5000 Generation Potential for Data Loss 19 Automatic generation of target reconfigurations -conforms to current environmental conditions. -analysis of tradeoffs and design decisions. -demonstration on real-world application!remote data mirroring.!runs in real time on a laptop. Future Work: -look at cost of reconfiguration. -apply approach to other problem domains. 20
Acknowledgements This work has been supported in part by NSF grants CCF-0541131, CNS-0551622, CCF-0750787, CNS-0751155, IIP-0700329, and CCF-0820220, Army Research Office W911NF-08-1-0495, Ford Motor Company, and a grant from Michigan State University s Quality Fund. 21 References [Fabregat05] R. Fabregat, Y. Donoso, B. Baran, F. Solano, and J. L. Marzo. Multi-objective optimization scheme for multicast flows: A survey, a model and a MOEA solution. In proceedings of the 3rd International IFIP/ACM Latin American Conference on Networking, pages 73-86, 2005 [Holland92] J. Holland. Adaptation in Natural and Artificial Systems. MIT Press, Cambridge, MA, USA, 1992. [Kephart03] J. O. Kephart and D. M. Chess. The Vision of Autonomic Computing. Computer, 36(1):41-50, 2003. [Lu07] J. Lu and W. Cheng. A -Algorithm-based Routing Optimization Scheme for Overlay Network. In Proceedings of the 3rd International Conference on Natural Computation, pages 421-425, 2007. [Montana02] D. Montana, T. Hussain, and T. Saxena. Adaptive Reconfiguration of Data Networks Using. In Proceedings of the and Evolutionary Computation Conference, pages 1141-1149, 2002. [Keeton04] K. Keeton, C. Santos, D. Beyer, J. Chase, and J. Wilkes. Designing for Disasters. In the Proceedings of the 3rd USENIX Conference on File and Storage Technologies, pages 59-62, Berkeley, CA, USA, 2004. [Keeton06] K. Keeton, D. Beyer, E. Brau, and A. Merchant. On the Road to Recovery: Restoring Data After Disasters. SIGOPS Operating Systems Review, 33(4):4-10, 2006. [Tseng06] S. Y. Tseng, Y. M. Huang, and C. C. Lin. Algorithm for Delay and Degree Constrained Multimedia Broadcasting on Overlay Networks. Computer Communications, 29(17):3625-3632, 2006. [Walsh04] W. E. Walsh, G. Tesauro, J. O. Kephart, and R. Das. Utility Functions in Autonomic Systems. In Proceedings of the First IEEE International Conference on Autonomic Computing, pages 70-77, 2004. [Wang08] D. Wang, J. Gan, and D. Wang. Heuristic Algorithm for Multicast Overlay Network Link Selection. In Proceedings of the Second International Conference on and Evolutionary Computing, pages 38-41, 2008. [Wilkes03] M. Ji, A. Veitch, and J. Wilkes. Seneca: Remote mirroring done write. In USENIX 2003 Annual Technical Conference, pages 253-268, Berkeley, CA. 22
Questions? Andres J. Ramirez ramir105@cse.msu.edu http://www.cse.msu.edu/~ramir105/ 23 Solution Space Optimal Solutions Acceptable Solutions All Possible Solutions 24
Setup First three experiments explore single dimensional concerns. -optimize only for cost, performance, or reliability -validate technique Fourth experiment explores multidimensional dynamic reconfiguration concerns. For every experiment, we conducted 30 trials. -account for stochastic nature. 25