Meta-heuristic Based Cloud Resource Provisioning Approach
Agenda for Today • Introduction • What is meta-heuristics? • How is it different from heuristics? • What makes them appropriate for resource provisioning as compared to heuristics? • Using a meta-heuristics approaches to solve the resource provisioning problem – Problem statement formulation and analysis of its complexity • What are the research question that are answered by meta-heuristics algorithms? • What are the open research questions in this area?
Introduction • The cloud computing paradigm is increasingly becoming mainstream and a growing number of companies and research organizations seek to gain value from its unique characteristics, service models, and deployment forms. • This gives rise to many different optimization problems both from the consumers’ and providers’ perspective.
Sample Problems • Cloud service providers need algorithms that can – reduce operational expenses and improve economies of scale to maximize profits and provide competitive prices in a highly competitive market. – algorithm that can predict for incoming request of VMs from the cloud users; – Algorithms that help improve utilization of the system – Efficient algorithms to optimize the current VM allocation.
Sample Problems • Cloud service users also need algorithms – prediction model to react to the dynamic price model of the CSPs – Algorithms that allow them to efficiently negotiate SLA with the Cloud service providers • Combined optimization issues – Algorithms for multi-criteria optimizations such as maximizing QoS while at the same time maximizing resource utilization
Agenda for Today • Introduction • What is meta-heuristics? • How is it different from heuristics? • What makes them appropriate for resource provisioning as compared to heuristics? • Using a meta-heuristics approaches to solve the resource provisioning problem – Problem statement formulation and analysis of its complexity • What are the research question that are answered by meta-heuristics algorithms? • What are the open research questions in this area?
What is a Meta-heuristic Method? • Meta-heuristic methods are one of the common strategies for solving NP-hard optimization problems • Meta-heuristic Heuristic + Randomization • We use the definition given by [Osman and Laporte 1996] – “ An iterative generation process which guides a subordinate heuristic by combining intelligently different concepts for exploring and exploiting the search space, learning strategies are used to structure information in order to find efficiently near-optimal solutions. “
Meta-heuristic Properties • Meta-heuristics properties – not problem-specific . – efficiently explore the search space to find (near-)optimal solutions. – approximate and usually non-deterministic.
Meta-heuristic Approaches • There are many meta-heuristic approaches that range from simple local search procedures to complex learning processes – simulated annealing (SA), [e.g., single solution approach] – evolutionary algorithms (EA), [e.g., population-based ] – ant colony optimization (ACO), [eg, of swarm intelligence] – particle swarm optimization(PSO) [e.g., population-based and swarm intelligence approaches] • Many different new variants are continually being proposed. • We will discuss Particle Swarm Optimization
Particle Swarm Optimization • Particle Swarm Optimization (PSO) [Kennedy and Eberhart 1995] algorithm – An adaptive method – Characterized by simplicity and effectiveness in wide range of application with low computational cost. – It also has fewer algorithm parameters than genetic algorithm. – PSO algorithm works well on most global optimal problems . • The above characteristics make PSO appealing to solve cloud resource allocation problem. – has become popular in solving resource allocation optimization problem
Particle Swarm Background • The particles are randomly placed in a problem search space. • A search area specifies an area that contains all possible solutions of a combinatorial problem. • Each particle in search space adjusts its “flying” pattern according to its own flying experience as well as the flying experience of other particles
Particle Attributes • Each particle will have – Velocity : The particle flying speed , which directs the flying of the particle. – Fitness : a fitness value (the objective function value), which will be evaluated by a fitness function to be optimised in each iteration. – Particle position (pbest) : Each particle knows its best position (this is the best fitness value so far reached by the particle) – Group position (gbest) : The best position so far among the entire group of particles (this is the best particle in terms of fitness in an entire population)
Particle Swarm Background • PSO is based on a set of flying particles that constitute a swarm move within the search space looking for the best solution. • Movement towards a promising area to get the global optimum
Velocity Update • Each particle adjusts its velocity (ie. travelling speed) dynamically corresponding to the flying experiences of itself and its neighbors pbest gbest velocity • At each time step, a particle moves toward its pBest and gBest.
Velocity Update..
Velocity and Position Updates
Illustration of PSO Process From Andry Pinto, et. al
Illustration of PSO Process From Andry Pinto, et. al
Illustration of PSO Process From Andry Pinto, et. al
Illustration of PSO Process From Andry Pinto, et. al
Illustration of PSO Process From Andry Pinto, et. al
Illustration of PSO Process From Andry Pinto, et. al
Particle Swarm Optimization Pseudo code Algorithm: PSO INPUT: N: number of tasks, I: number of iterations OUTPUT: 1. D[]=N //particle dimension 2. Initialize position randomly 3. Initialize velocity randomly 4. REPEAT 5. Evaluate fitness of individual particle. 6. IF fitness is better than pbest 7. pbest = fitness ; 8. END 9. gbest = neighboughr (pbest); 10. Modify velocities based on personal best and global best. 11. Modify position based on current location and velocity. 12. UNTIL Terminate on some condition. 13. Return result 14. END algorithm
Particle Swarm Optimization Initialization step The set of particles # of iterations F(p) F(p) F(p) Convergence Update velocity, position, gbest, pbest
Agenda for Today • Introduction • What is meta-heuristics? • How is it different from heuristics? • What makes them appropriate for resource provisioning as compared to heuristics? • Meta-heuristics Cloud resource provisioning approach – Problem statement formulation and analysis of its complexity • What are the research question that are answered by meta-heuristics algorithms? • What are the open research questions in this area?
Preliminaries
Problem Formulation
Illustration 1 0 0 1 0 0 0 0 1 0 1 0 0 0 1 1 0 0 0 0 1 0 1 0 1 0 0 • Which task to machine assignment is the best?
NP Hardness • Even though this problem has been intensively investigated, exact polynomial algorithms have not been found yet. • Furthermore, we can verify that the problem is NP- hard by reduction to the partition set problem, that, even for n = 2, this problem is NP-hard. • Meta-heuristic optimization algorithm can be used to solve NP-hard problems.
Meta-heuristic Resource Allocation • Application of meta-heuristic techniques to Cloud Resource Allocation League Championship Algorithm (LCA) MalaKalraaSarbjeetSingh, A review of metaheuristic scheduling techniques in cloud computing, Egyptian Informatics Journal, Volume 16, Issue 3, November 2015, Pages 275-295
Cloud Resource Allocation 1 0 0 1 0 0 0 0 1 0 1 0 0 0 1 1 0 0 0 0 1 0 1 0 1 0 0
Applying PSO to Resource Provisioning Task to resource mapping instance Particle in PSO 4 5 6 1 2 3 7 8 A representation of task to resource mapping as a PSO particle • Each position in the particle represents a task and the value at that position represents the mapping of the task to a resource. • Thus the particle represents mapping of resource to a task.
Open Problems • Research question 1 : Develop a hybrid algorithm that combines metaheuristic techniques and classical Artificial Intelligence (AI) and Operation Research (OR) methods such as greedy algorithm, backtracking techniques, beam search or constrained programming for cloud data center resource management. • Research question 2 : Investigation and development of a hybrid approach that combines single solution approach (eg., simulated annealing (SA) and population-based metaheuristic technique for cloud data center resource management. • Research question 3 : Metaheuristic algorithm that can (i) predict for incoming request of VMs from the cloud users; (ii) prediction model to react to the dynamic price model of the CSPs
Recommend
More recommend
