Placement (Scheduling) • Optimal mapping of VMs – – to physical hosts in a data center (cloud) across multiple clouds • • • • When? – • Federation and bursting Multi-cloud service deployment Third-party broker scenarios Admission of new service, upon elasticity, hardware failure, periodically Optimal? – Service Provider perspective • – Performance (hosts, VMs), cost, guarantees, nonfunctional criteria (location, isolation, trust, risk, ecoefficiency, etc.) Infrastructure Provider perspective • Provisioning cost, consolidation, isolation, SLA violations, etc. 1 Placement (cont.) • Further considerations – – – – – – – – – Historical performance data Benchmarking and application profiling Co-location and (anti)affinity End-user location Data constraints (legislations) Federation (lack of control over remote resources) Dynamicity - providers, prices, performance, workloads, etc. change over time (Live) Migration overhead (end-user) SLAs – perspectives 1. 2. 3. All management actions are SLA-driven Placement = SLA refinement SLAs are just another criteria 2 Sense Example Approach • Act Plan Combinatorial optimization formulations Packing formulations for data centers (MMKP) – • • • 0-1 integer programming (assignment problems) for multi and federated clouds – • • – Multi-dimensional (CPU, memory, disk, network), multi-choice (many physical hosts) Knapsack Problems Policies for load balancing, power saving (consolidation), SLA protection Scalability improvements (fractional 2-approximation) Optimize service performance and/or cost, with service layout (load balancing), budget, VM configuration, etc. as constraints. Model uncertainty (changing conditions in providers, offers, performance, etc.) and migration overhead Approximations (greedy heuristics) for scalability Placement - Experiences • Reservoir (and IBM SUR grant) – • OPTIMIS – • Bursting and Federated/Multi-cloud deployment based on functional and non-functional criteria (trust, risk, eco-efficiency, cost) Vision Cloud – • Placement optimization within clouds and across federated clouds. SLA protection and/or load balancing, consolidation, revenue maximization Placement of compute close to data Various Grid research projects – QoS, SLA management, advance reservations, co-allocation, fair-share scheduling, job management, performance predictions, etc 4 Outlook and perspectives • Placement of services (that use compute, data, and networking) – Compute, data, and/or network intense – Network aware vs. managed networks • Holistic view of placement problems for all cloud architectures • Interactions with related problems – Time perspective (short - long) • Placement and admission control – Abstraction level (low - high) • Placement and governance Selected references • D. Breitgand, A. Marashini, and J. Tordsson. Policy-Driven Service Placement Optimization in Federated Clouds, IBM Haifa Labs technical report H-0299, 2011 • J. Tordsson, R.S. Montero, R.M. Vozmediano, and I.M. Llorente. Cloud brokering mechanisms for optimized placement of virtual machines across multiple providers. Future Generation Computer Systems, 2011, accepted. • B. Rochwerger, J. Tordsson, C. Ragusa, D. Breitgand, S. Clayman, A. Epstein, D. Hadas, E. Levy, I. Loy, A. Maraschini, P. Massonet, H. Munoz, K. Nagin, G. Toffetti, and M. Villari. Reservoir - when one cloud is not enough, IEEE Computer 2011, accepted. • W. Li, J. Tordsson, and E. Elmroth. Modelling for Dynamic Cloud Scheduling via Migration of Virtual Machines (tentative), in preparation, 2011 • P-O Östberg, Virtual Infrastructures for Computational Science, PhD thesis, 2011 • J. Tordsson. Portable Tools for Interoperable Grids, PhD thesis, 2009 6