Cost-aware real-time divisible loads scheduling in cloud computing
Loading...
Date
2022
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Teknologi Malaysia
Abstract
Cloud computing has become an indispensable alternative for processing big data in science, engineering and analytics. Today, cloud service providers typically offer users virtual machines with different combinations of configurations and prices. Since each user has different preferences and priorities, the problem of allocating the minimum-cost processors while meeting a deadline becomes increasingly complex. Moreover, most previous research have assumed that processors in cloud computing are homogeneous. However, in reality, it consists of heterogeneous processors with different speeds. This thesis examined user preference adaptation for scheduling divisible workloads in a cloud computing platform with deadline constraint as a quality of service (QoS) criteria. The workload allocation approach used in this research is Real-time Divisible Load Theory (RT-DLT). Two crucial problems were investigated: choosing the minimum cost resource combination when processors are heterogeneous in terms of speed and allowing cloud users to have different preferences in terms of their cost while being able to meet their specified deadline. For the first problem, an algorithm called Cost Aware Real-Time Divisible Load Theory (CARTDLT) was developed and a Worker Selection Strategy (WSS) was introduced into the RT-DLT scheduling framework. The additional strategy is mainly to select the best combination of processing nodes that results in the desired total cost based on user requirements. For the second problem, an algorithm called Market-Oriented Real-time Divisible Load Theory (MORTDLT) was developed to group the resources and optimally distribute the load fragments among the available resources according to the user's preference. The proposed algorithm was evaluated through the experimental evaluation using MATLAB and CloudSim 3.0.3, and compared with Min-Min, Max- Min and Sufferage algorithms. The CARTDLT algorithm showed a cost improvement of 45.06% over the Max-Min algorithm, 42.52% over the Min-Min algorithm and 45.57% over the Sufferage algorithm. For low priority user preference, the MORTDLT algorithm showed a cost improvement of 61.22 % over the Max-Min algorithm, 83.58 % over the Min-min algorithm and 49.76 % over the Sufferage algorithm. For high priority user preference, the MORTDLT algorithm showed a cost improvement of 54.49% over the Max-min algorithm, 43.12% over the Min-Min algorithm and 57.60% over Sufferage algorithm. The results indicate that the proposed algorithms have the lowest computational cost and imbalance, and ensure compliance with the deadline without compromising other performance metrics such as makespan. The CARTDLT and MORTDLT algorithms will help cloud users to have different preferences on their costs while meeting the deadline for their job.
Description
Thesis (Doctor of Philosophy)
Keywords
Cloud computing, Electronic data processing—Distributed processing