QoS Resource Management for Cloud Federations Gaetano F. Anastasi National Council of Research (CNR), Pisa, Italy Pisa, June 16th, 2014 gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 1/13 1 QoS Resource Management 2 Conclusion and Future work gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 2/13 Contrail Resource Mgmt. • Contrail Federation sits in the middle - serving the users - exploit efficiently the providers • Must operate a trade-off between two apparent discording objectives - optimize for users (e.g., minimize cost) - optimize for providers (e.g., ensuring fairness in providers’ revenues) • Federation is an autonomous entity - directly corresponds for providers’ violations gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 3/13 Allocation Problem (1/5) Common Cloud Approach • bin-packing • performance corrective actions (e.g. VM migration host-basis) Federation Issue • migration across providers is very expensive - increasing down-time (provider negotiation, image transfer time, etc.) • smart allocation to avoid migration gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 4/13 Allocation Problem (2/5) Problem Formulation • Allocating N services on M providers • Respecting QoS constraints from user (e.g. dedicated bandwidth, bounded response time, provider location, etc.) • Input: app reqs and provider characteristics • Objectives - Minimizing user criteria (e.g. cost) - Maximize federation revenue (minimizing risk of paying penalties) gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 5/13 Allocation Problem (3/5) Broker case • Sub-problem of the federation allocation • Optimization on cost • Loosely-coupled architecture (cannot provide additional guarantees) • Can mitigate vendor lock-in gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 6/13 Allocation Problem (4/5) QBrokage • Genetic approach Table: Lock-in Degree • QoS = cost, ram, storage, location • Support multiple provider cost models • Flexible to plug-in • Scalable (250 ms increment on m-time from 50 to 500 P) gaetano.anastasi@isti.cnr.it QBrokage Naive P Cost (P) Cost (P) 50 100 150 200 0.082 (1.5) 0.078 (2.35) 0.082 (3.20) 0.075 (3.50) 0.082 (1.0) 0.077 (1.0) 0.080 (1.0) 0.073 (1.0) QoS Management for Cloud Federations 7/13 Allocation Problem (5/5) Addressing Federation Issues • Resource cost model that consider providers’ reputation - federation directly corresponds penalties - acquiring resources from bad providers may cost more at the end • Monitoring applications and SLA violations - updating reputation and costs per provider • Dynamic provider reputation - transient overloads must be considered - avoiding fluctuations gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 8/13 1 QoS Resource Management 2 Conclusion and Future work gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 9/13 PaaS Federation (1/2) Today • libcloud drivers for the Contrail Federation • PaaS layers can run transparently on top of the federation Tomorrow • Towards a PaaS Federation • Third-party layers can benefit from quantity and quality of resources • Reusing existing federated mechanisms gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 10/13 PaaS Federation (2/2) Translating PaaS requirements • To address heterogeneity of providers • Enforce high-level QoS guarantees Enforcement strategies • VM Elasticity - Increasing/decreasing VM number - Less precise but widely supported • Resource scaling - Increasing virtual resource quantity - Require underlying support gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 11/13 Conclusion • Inter-Cloud Computing brings benefits in terms of interoperability and dependability • ...but the complexity of management is increased and poses many challenges • Cloud Federations cannot scale if relying on ad-hoc methods • Leveraging other research fields - Cloud Control - Real-time techniques - Cognitive heuristics gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 12/13 Contribution Gaetano F. Anastasi, Emanuele Carlini and Patrizio Dazzi Smart cloud federation simulations with CloudSim In Proceedings of the first ACM workshop on Optimization techniques for resources management in clouds, ORMaCloud ’13 , June 2013 https://github.com/ecarlini/smartfed/ Gaetano F. Anastasi, Emanuele Carlini, Massimo Coppola, Patrizio Dazzi, Aliaksandr Lazouski, Fabio Martinelli, Gaetano Mancini and Paolo Mori Usage Control in Cloud Federations In Proceedings of the IEEE International Conference on Cloud Engineering, IC2E 2014 , March 2014 Gaetano F. Anastasi, Emanuele Carlini, Massimo Coppola and Patrizio Dazzi QBROKAGE: A Genetic Approach for QoS Cloud Brokering In Proceedings of the 7th IEEE International Conference on Cloud, CLOUD 2014 (To appear), June 2014. ´ Patrizio Dazzi, Alberto Gotta, Matteo Mordacchini, Andrea Gaetano F. Anastasi, Pietro Cassara, Passarella A Hybrid Cross-Entropy Cognitive-based Algorithm for Resource Allocation in Cloud Environments In Proceedings of the 8th IEEE International Conference on Self-Adaptive and Self-Organizing Systems, SASO 2014 (Under revision), June 2014. Gaetano F. Anastasi, Emanuele Carlini, Massimo Coppola, Patrizio Dazzi and Marco Distefano An OVF Toolkit Supporting Inter-Cloud Application Splitting In Proceedings of the IEEE International Conference on Cloud Networking, Cloudnet 2014 (Under revision), June 2014. gaetano.anastasi@isti.cnr.it QoS Management for Cloud Federations 13/13