Application of Soft-Computing Techniques to the Design of Meta-Scheduling Systems for Grid Computing M. Sc. Rocío Pérez de Prado Telecommunication Engineering Department University of Jaén. Spain Dortmunder Regelungstechnische Kolloquien Lehrstuhl für Regelungssystemtechnik Technische Universität Dortmund. Germany October 28, 2010
Outline 1 Introduction 2 Fundamentals in grid scheduling Grid Scheduling Structure Scheduling strategies and challenges Soft-computing techniques 3 4
Grid Systems Grid Large scale, heterogeneous and autonomous systems Geographically distributed Interconnected with high speed networks Grid computing New paradigm of distributed computing Large scale computational problems
Grid computing applications Grid computing in science, engineering and technology Bioinformatics Medical image analysis High energy physics
Introduction Grid computing applications (II) Grid Computing as a new business model: utility grid University of Jaén Dortmunder Regelungstechnische Kolloquien 28.10.2010
Grid Systems advantages & challenges Ability to build dynamic applications that use distributed resources in order to improve a tness function. Utilization of resources that are located in a particular domain to increase throughput or reduce execution costs. Adaptation of parallel programs. Eectively handling dynamic and heterogeneous grid resources. Grid scheduling. Grid Scheduler Manager of the workow Resource discovery and publishing Submission and monitoring jobs Objectives Throughput, Resource utilization,turn around times, Response Times
Fundamentals in grid scheduling Grid Scheduling Structure Scheduling strategies and challenges Soft-computing techniques Fundamentals in grid scheduling Types of grid scheduling Computational, Independent, Hierarchical (meta-scheduling). Scheduling within Grid architecture Virtualization Multi-objective Performance criteria Optimization criteria
Grid scheduling structure: two level-grid Fundamentals in grid scheduling Grid Scheduling Structure Scheduling strategies and challenges Soft-computing techniques
Scheduling strategies and challenges Fundamentals in grid scheduling Grid Scheduling Structure Scheduling strategies and challenges Soft-computing techniques Queue-based strategies EASY Backlling or EDF (Condor, Grid Service Broker) Schedule-based strategies Certain level of QoS must concern grid state Adaptive scheduling: future and present grid state to avoid or prevent performance deterioration Heterogeneous and dynamic environment Resources computational capacity and availability change with time: fully dynamic environment with uncertainty Resources fall down, become reserved, change access policies or join the system Self-adapting and exible schemas
Soft-Computing techniques Fundamentals in grid scheduling Grid Scheduling Structure Scheduling strategies and challenges Soft-computing techniques LÓGICA BORROSA Sistemas Borroso AlgoritmosGenéticos Evolutivo Borrosos Redes Neuronales Sistemas Borrosas Borroso Neuronales Redes Neuronales Genéticas REDES NEURONALES Soft Computing COMPUTACIÓN EVOLUTIVA Redes Bayesiano geneticas RAZONAMIENTO PROBABILISTICO
Fundamentals in grid scheduling Grid Scheduling Structure Scheduling strategies and challenges Soft-computing techniques Fuzzy Logic and Evolutionary Computation Fuzzy Rule-Based Systems Expert systems IF-THEN fuzzy rules and fuzzy sets Diverse areas: intelligent control of elevator systems, classication in speech/music discrimination applications... Scheduling: uncertainty in resources state information Quality of theirs knowledge bases-learning system Classical learning strategies New learning strategies
General objectives of our research 1 Study and analysis of fuzzy expert systems to design scheduling systems for Grid computing Grid state knowledge Dynamism and uncertainty Fuzzy Logic 2 Application of knowledge acquisition mechanisms to expert schedulers Quality of expert knowledge and acquisition process features
Specic objectives of our research 1 Design of an expert meta-scheduler based on fuzzy systems for its application to Grid computing 1 Specication of fundamental components of the expert system: fuzzication, inference and defuzzication systems 2 Grid state featuring: resource domains state variables denition 3 Design of associated data bases 2 Integration knowledge acquisition mechanisms to the expert scheduler 1 Application of classical learning strategies in fuzzy rule-based systems 2 Study of soft-computing techniques to improve schedulers design and properties 3 Design of alternative techniques to reduce optimization time
Grid scenarios GridSim-based toolkit based on real world settings. Network conguration, resources, reservation, maintenance and workload traces Czech National Grid Infrastructure Metacentrum project Grid scenario and workload traces: Czech National Grid Infrastructure Metacentrum CESNET project (operator of academic network of the Czech Republic -National Research and Education Network, NREN) 14 heterogeneous RDs-210 machines, 806 CPUs running Linux AuverGrid Traces from Grid Workload Archive AuverGrid. Production grid platform: 5 heterogeneous RDs EGEE project (Enabling Grids for E-science in Europe) LCG (Large hadron collider Computing Grid project) middleware
Degree of membership Degree of membership Degree of membership Degree of membership 0. 8 0. 6 0. 4 0. 2 1 0. 8 0. 6 0. 4 BAJO MEDIO ALTO 1 0 BAJO MEDIO ALTO 0 0.1 0. 2 0. 3 0.4 0.5 0.6 0.7 0.8 0.9 1 RECURSOS 0. 2 0 0 0.1 0. 2 0. 3 0.4 0.5 0.6 0.7 0.8 0.9 1 BAJO MEDIO ALTO Degree of membership Degree of membership Degree of membership 1 0. 8 0. 6 0. 4 RECURSOS 0. 2 0 BAJO 0 0.1 0. 2 0. 3 0.4 MEDIO 0.5 0.6 0.7 0.8 0.9 ALTO1 1 0. 8 0. 6 0. 4 RECURSOS 0. 2 0 0 0.1 0. 2 0. 3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 0. 8 0. 6 0. 4 BAJO RECURSOS MEDIO ALTO 0. 2 0 BAJO 0 0.1 0. 2 0. 3 0.4 MEDIO 0.5 0.6 0.7 0.8 0.9 ALTO1 1 0. 8 0. 6 0. 4 RECURSOS 0. 2 0 0 0.1 0. 2 0. 3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 0. 8 0. 6 0. 4 RECURSOS BAJO MEDIO ALTO 0. 2 0 0 0.1 0. 2 0. 3 0.4 0.5 0.6 0.7 0.8 0.9 1 RECURSOS Zd dd dd dd Zd dd dd dd ZE Ed Ed Ed Zd Zd ZE dd dd de de dd dd de de Ed Ed Ee Ee d d E MUYBAJO BAJO MEDIO ALTO MUYALTO 1 0.8 0.6 0.4 0.2 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 d d E SALIDA Introduction Fuzzy Rule-Based Meta-Scheduler Structure Adaptation of fuzzy systems to grid meta-scheduling for grid computing Degree of membership
Grid state featuring Meta-scheduler input features Number of free processing elements (FPE): Number of free processing element within RDi. Previous Tardiness (PT): Sum of tardiness of all finished jobs in RDi. Resource Makespan (RM): Current makespan for RDi. Resource Tardiness (RT): Current tardiness of jobs within RDi. Previous Score (PS): Previous deadline score of already finished jobs in RDi. Resource Score (RS): Number of non delayed jobs so far in RDi. Resources In Execution (RE): Number of resources currently executing jobs within RDi. RD Selection Factor FPE LOW MIDDLE HIGH PT 1 RM RT PS RS 0.8 RE VERY-LOW LOW MIDDLE HIGH VERYHIGH 1 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Learning challenge Rules evolution INITIAL RULES V2 A V3 B V8 E V8 G R1: V4 H V3 C V6 D V6 F R2: V8 I CLASSIC CROSSOVER SOFT MUTATION V5 R1 : V2 A V3 B V6 D V8 F V8 I a 1 b 1 c 1 d 1 R2 : V3 C V8 E V9 G V4 H V5 UNIFORM CROSSOVER a 0 b 0 c 0 d 0 R1 : V3 B V8 F V9 G V8 I HARD MUTATION V2 R2 : V2 A V3 C V6 D V8 E V4 H a 0 b 0 c 0 d 0 Rules encoding R i = if ω 1 is A1nand/or...ω m is Amn then y is Bn : w i (1) R i : [ a1... am bn cn w i ] (2)
Introduction Application of classical genetic learning strategies to expert scheduling systems Learning strategy based on Michigan approach Learning strategy based on Pittsburgh approach Prado, R.P., García-Galán, S., Yuste A.J., Muñoz-Expósito, J.E. Genetic Fuzzy Rule-Based Scheduling System for Grid Computing in Virtual Organizations. Soft Computing. Springer. Accepted. doi: 10.1007/s00500-010-0660-5, September 2010. In press. Prado, R.P., García-Galán, S.,Yuste, A.J., Muñoz Expósito, J.E., Sánchez A.J., Bruque, S. Evolutionary Fuzzy Scheduler for Grid Computing. 10th INTERNATIONAL WORK-CONFERENCE ON ARTIFICIAL NEURAL NETWORKS (IWANN 2009), June 2009, Salamanca (Spain)
Z Introduction Improving genetic strategies to their application to expert scheduling systems Hybrid Pittsburgh-Michigan structure Prado, R.P., García-Galán, S.,Yuste, A.J. Yuste, Muñoz Expósito, J.E.. A fuzzy rule-based meta scheduler with evolutionary learning for grid computing. Engineering applications of articial intelligence. 23 (7), 1072 1082. doi:10.1016/ j.engappai. 2010. 07.002, October 2010
& / Introduction Improving genetic strategies to their application to expert scheduling systems (II) Hybrid Pittsburgh and Cooperative/Competitive structure Prado, R.P., García-Galán, S., Muñoz Expósito, J.E., Yuste, A.J., Bruque, S. Genetic Fuzzy Rule-Based Meta-Scheduler for Grid Computing. FOURTH INTERNATIONAL WORKSHOP ON GENETIC AND EVOLUTIONARY FUZZY SYSTEMS (GEFS 2010). March 2010. Mieres, Asturias (Spain)
Application of other evolutionary techniques to the learning of scheduling expert systems- Dierential Evolution Prado, R.P., García-Galán, S., Muñoz Expósito, J.E.,Yuste, A.J. and Bruque, S.Learning of fuzzy rule-based meta-schedulers for grid computing with dierential evolution. INTERNATIONAL CONFERENCE ON INFORMATION PROCESSING AND MANAGEMENT OF UNCERTAINTY IN KNOWLEDGE-BASED SYSTEMS (IPMU 2010). June July, 2010. Dortmund (Germany).
Adaptation of swarm intelligence to knowledge acquisition in fuzzy rule-based systems and its application to Grid computing schedulers Knowledge Acquisition with a Swarm Intelligence Approach (KASIA) Prado, R.P., García-Galán, S., Muñoz Expósito, J.E.,Yuste, A.J. Delgado. Knowledge Acquisition in Fuzzy Rule Based Systems with Particle Swarm Optimization. IEEE Transactions on fuzzy systems. Accepted. August 2010. doi: 10.1109/ TFUZZ. 2010. 2062525. August 2010. In press.
Applying multi-objective techniques in knowledge acquisition for fuzzy rule-based schedulers Lehrstuhl für Regelungssystemtechnik (RST). Fakultät für Elektro- und Informationstechnik.Technische Universität Dortmund. Germany Multi-objective general approach Conicting parameters optimization. General Pareto theory Makespan min S i Sched{max j J T j } (3) Flowtime min S i Sched{ T j } (4) j J T j nalization time for J j Sched all the possible schedules J set of considered jobs
Main results Fuzzy rule-based meta-scheduler training 1.78 x 106 Convergence speed (78-64 iterations) Reduction of computational eort (252 rule bases evaluations) Accuracy improvement Fitness (s) 1.76 1.74 1.72 1.7 1.68 1.66 1.64 KASIA PITTSBURGH 1.62 (2.23%) 0 10 20 30 40 50 60 70 80 90 100 Iteration Fitness Max Min Average Standard D Condence I (95%) KASIA 1,654,505.4 1,553,820.1 1,630,479.5 36,778.7 1,617,318.41, 1,643,640.50 GA-Pitts 1,684,235.5 1,625,058.2 1,667,586.2 19,757.8 1,660,515.96, 1,674,656.43
Main results Fuzzy rule-based meta-scheduler training Multi-objective learning Deeper exploration of search space Trade-o among conicting criteria 7.05 x 104 7 Evolution of Average Non Dominated Solutions MO-KASIA MO-Pittsburgh 6.95 Flowtime (s) 6.9 6.85 6.8 6.75 1.7 1.72 1.74 1.76 1.78 1.8 1.82 1.84 Makespan (s) x 10 6
Validation and comparison with other classical scheduling strategies Metric/Strategy Fuzzy-KASIA Fuzzy-Pittsburgh EASY-BF ESG+LS periodical Makespan (s) 1,633,719.8 1,659,926.3 1,749,586.0 1,973,151.4 Flowtime (s) 87,765.539 88,865.534 87,491.471 83,379.182 Weighted usage (%) 48.184 46.420 44.471 34.744 Classic usage (%) 59.360 56.528 47.013 40.914 Tardiness (s) 4,683.250 4,757.286 3,235.311 1,274.822 Slowdown (s) 197.036 192.049 184.352 17.522 Accuracy improvement of KASIA vs. genetic strategy: 1.58%, training index KASIA outperforms machine usage: classical y weighted usage by 4.77% y 3.66% Makespan improvement: EASY-BF by 6.62% and ESG+LS periodical by 17.20% Usage machine improvement: weighted and classical usage 7.71% and 20.81%
Number of waiting/running jobs: KASIA, EASY-BF and ESG+LS periodical KASIA EASY-BF ESG+LS periodical
Cluster usage: KASIA, EASY-BF and ESG+LS periodical KASIA EASY-BF ESG+LS periodical
Conclusions Mechanism for ecient scheduling in Grid computing Adaptive scheduling strategies Fuzzy rule-based systems Adaptability to environment changes Ability to model uncertainty in Grid system state Scheduling: Improvement of other classical scheduling strategies results Knowledge Acquisition Improvement in convergence behaviour, accuracy and/or computational cost vs. traditional learning schemas in fuzzy rule-based systems
Publications International Journals Adaptation of swarm intelligence to knowledge acquisition in fuzzy rule-based systems and its application to Grid computing schedulers Prado, R.P., García-Galán, S., Muñoz Expósito, J.E.,Yuste, A.J.. Knowledge Acquisition in Fuzzy Rule Based Systems with Particle Swarm Optimization. IEEE Transactions on fuzzy systems. Accepted. doi: 10.1109/ TFUZZ. 2010. 2062525. August 2010. In press. Application of classical genetic learning strategies to expert scheduling systems Prado, R.P., García-Galán, S., Yuste, A.J., Muñoz Expósito, J. E. Genetic Fuzzy Rule-Based Scheduling System for Grid Computing in Virtual Organizations. Soft Computing. Springer. Accepted. doi: 10.1007/s00500-010-0660-5, September 2010. In press. Improving genetic strategies to their application to expert scheduling systems Prado, R.P., García-Galán, S.,Yuste, A.J. Delgado, Muñoz Expósito, J.E.. A fuzzy rule-based meta scheduler with evolutionary learning for grid computing. Engineering applications of articial intelligence. 23 (7), 1072 1082. doi:10.1016/ j.engappai. 2010. 07.002, October 2010.
Publications International Conferences Application of other evolutionary techniques to the learning of scheduling expert systems- Dierential Evolution Prado, R.P., García-Galán, S., Muñoz Expósito, J.E.,Yuste, A.J. and Bruque, S.. Learning of fuzzy rule-based meta-schedulers for grid computing with dierential evolution. INTERNATIONAL CONFERENCE ON INFORMATION PROCESSING AND MANAGEMENT OF UNCERTAINTY IN KNOWLEDGE-BASED SYSTEMS (IPMU 2010). June July, 2010. Dortmund (Germany). Improving genetic strategies to their application to expert scheduling systems Prado, R.P., García-Galán, S., Muñoz Expósito, J.E., Yuste, A.J., Bruque, S. Genetic Fuzzy Rule-Based Meta-Scheduler for Grid Computing. FOURTH INTERNATIONAL WORKSHOP ON GENETIC AND EVOLUTIONARY FUZZY SYSTEMS (GEFS 2010). March 2010. Mieres, Asturias (Spain). Design of alternative techniques to reduce optimization time Parra, F., García-Galán, S.,Yuste, A.J., Prado, R.P., Muñoz Expósito, J.E.. A Method to Minimize Distributed PSO Algorithm Execution Time in Grid Computer Environment. 3rd. INTERNATIONAL WORK-CONFERENCE on the INTERPLAY between NATURAL and ARTIFICIAL COMPUTATION. June, 2009. Santiago de Compostela (Spain). Application of classical genetic learning strategies to expert scheduling systems Prado, R.P., García-Galán, S.,Yuste, A.J., Muñoz Expósito, J.E., Sánchez, A.J., Bruque, S. Evolutionary Fuzzy Scheduler for Grid Computing. 10th INTERNATIONAL WORK-CONFERENCE ON ARTIFICIAL NEURAL NETWORKS (IWANN 2009), June 2009, Salamanca (Spain).
Current works Improvement of suggested knowledge acquisition for fuzzy rule-based meta-schedulers based on swarm intelligence Knowledge Acquisition with a Swarm Intelligence Approach Automatic selection of expert scheduler knowledge bases Design of decentralized meta-schedulers for virtual organizations Lehrstuhl für Regelungssystemtechnik (RST). Fakultät für Elektro- und Informationstechnik.Technische Universität Dortmund. Germany Design of new scheduling strategies based on gaussian mixture models and evolutionary techniques