Multicriteria based Resource Allocation Policies for Cloud based Systems

Abstract

Modern distributed computing systems provide ubiquitous services as utilities to end-user. Major advantages of such systems are characteristics such as ease of use, flexibility, and scalability. These systems are often associated with variable traffic patterns and large scale. The objective of service providers is to earn profit by deploying minimum resources while that for end-user is to use resources for minimum time. Conflicting objectives make the problem of resource allocation in these systems NP complete thereby, creating a demand for efficient resource allocation policies. Heterogeneity in architectures and end user bases ask for multiple criteria or joint resource allocation policies that optimize multiple aspects of resource allocation. This thesis addresses some of the problems associated with resource allocation in these advanced distributed systems. Resource allocation policies are presented for cloud and fog cloud systems that optimize response time and energy consumption from the perspective of the end user. For cloud computing systems, the major concern is end user response time. Reduction in response time not only helps service providers to attract more customers but also benefits the end-user with improved cost and quality of service. In order to reduce the response time for end-user in cloud, two resource allocation policies are proposed: (i.) greedy allocation, and (ii.) ACOAHP based allocation. The greedy resource allocation policy is based on Paretooptimal joint allocation (POJA) of compute and network resources. The ACOAHP based resource allocation approach reduces the overall end user response time by Nature inspired joint allocation (NIJA) of compute and network resources. Mathematically and experimentally it is verified that the proposed approaches report superior results in comparison to existing resource allocation policies for cloud systems. For the fog cloud hybrid architecture that operate at the edge of the network, the major concern is delay sensitive execution. Two AHP based resource allocation policies are presented for the fog cloud hybrid systems with the aim to reduce the response time for end-user. These policies differ in the way they assign weights to the multiple criteria. The first policy uses the principal Eigenvector to derive predefined criteria weights. However, the second policy finds the weights of the criteria dynamically from the data using the AHP technique of Simultaneous Evaluation of Criteria and Alternatives (SECA). Experimentally, it is verified that the proposed approaches outperform state of the art resource allocation policies for fog cloud hybrid architectures. In fog cloud systems, apart from response time, another major concern is the limited battery life of the end user mobile devices. A noisy channel affects the energy consumption of end-user devices due to attenuation and distortion in the transmitted signal. The attenuation and distortion can be attributed to the multipath propagation and mobility of end-users. In order to address the issue of power consumption of end-user devices, an energy efficient resource allocation policy is presented based on Markov Decision Process. The Markovian offloading approach considers a Rayleigh fading time varying network between the end-users and the fog cloud servers. Experimentally, it is shown that the proposed approach is able to extend the battery life of end-user devices.