Optimizing resource allocation using multi-objective particle swarm optimization in cloud computing systems

Abstract

Allocating resources in data centers is a complex task due to their increase in size, complexity, and consumption of power. At the same time, consumers' requirements regarding execution time and cost have become more sophisticated and demanding. These requirements often conflict with the objectives of cloud providers. Set against this background, this thesis presents a model of resource allocation in cloud computing environments that focuses on developing the allocation process in three phases: (i) negotiation between consumers and providers to select the data center, (ii) scheduling tasks inside data centers, and (iii) scheduling virtual machines (VMs) to physical machines. The proposed model attempts to optimize each phase by applying multi-objective optimization (MOO) and many-objective optimization (MaOO) using a particle swarm optimization (PSO) algorithm.

In more detail, a parallel PSO (PPSO) algorithm based on multi-objective was therefore developed to improve the SLA negotiation process between consumers and providers. The main insight of this algorithm is that SLA negotiation can be automated and the PSO can be parallelized to minimize negotiation time and to maximize system throughput, thus increasing the profits of providers.

A many-objective PSO (MaOPSO) algorithm based on a modified ranking strategy was developed to improve the task scheduling problem in each data center. The novelty of this algorithm lies in using a modified ranking strategy to minimize evaluation time and improve the quality of the results. The algorithm was executed within the constraints of the tight deadline to improve performance in terms of both waiting time and completion time.

Finally, VM allocation was improved by applying a many-objective PSO to allocate VMs in physical machines after clustering the hosts. Here the novelty lies in applying PSO and K-means when clustering hosts to improve VM allocation and migration, thus maximizing resource utilization and performance whilst reducing power consumption.
Most notably, SLA Negotiation reduced waiting time and completed time by up to 20%. Additionally, it increased the throughput by about 20%. The proposed SLA negotiation reduced the rates of SLA violations by about 25%. On the other hand, the proposed MaOPSO task algorithm reduced the waiting time and completed time by 15% and 20% respectively. It increased the throughput up to 15% and the profits up to 15%.

With respect to MaOPSO VM allocation, it improved resource utilization by up to 20%. Additionally, it reduced the power consumption by 25% compared to other algorithms. Profits are indirectly increased by improving utilization up to 20%. Finally, the MaOPSO VM algorithm led to an increased throughput of 20%, a reduced waiting time of 15%, and reduced the completed time up to 15%.

More information

Identifiers

Catalogue record

Export record