Developing a metaheuristic solution model to task scheduling problems in cloud systems

Cloud computing technology is the sum of virtualizable and scalable resources that enable data to be hosted and evaluated on the internet, and users to pay only for the resources they consume. Many reasons, such as the development of the Internet infrastructure, the spread of the Internet of Things technology, the rapid growth of big data and the emergence of studies on this, and the developments in artificial intelligence studies, have led to the widespread use of cloud technologies. One of the most important mechanisms of cloud computing is virtual machines. Virtual machines are created from resources on the cloud system in line with the needs of customers. Depending on the business volume, the number of virtual machines created by customers and their features may vary. Customers pay certain fees to the cloud provider based on the features and duration of use of these virtual machines. Incorrect scheduling of tasks that need to be run on virtual machines leads to increased task completion time (makespan), and naturally, increased cost for the customer. This increase in task completion time indirectly affects the energy consumption and maintenance costs of the cloud provider. For this reason, the use of a good task scheduling algorithm in cloud systems is mandatory for both the customer and the cloud provider. Task scheduling is an NP-hard type problem. The use of metaheuristic algorithms instead of deterministic approaches is often preferred in solving such problems in terms of performance. However, due to the type of problem, metaheuristic algorithms based on random search may get stuck in local minimums. This possibility may increase in case where the number of tasks and virtual machines has increased. For this reason, the metaheuristic algorithms preferred should use efficient mechanisms to overcome this problem. This study proposes a solution based on Jellyfish Search Optimizer, one of the current metaheuristic algorithms, which uses a different approach mechanism to solve this problem. The most unique aspect of the proposed method is that it allows dynamic population growth with a different similarity control to get rid of local minimums more quickly. Thus, a more efficient exploration process is realized in the search space. In addition, the multi-thread and multi-process analyses of this algorithm for the task scheduling problem were also made within the scope of this thesis. The performance of the proposed method has been comparatively tested and proven for different scenarios in the CloudSim simulator.