Genetic Programming Hyper-Heuristic for Cost and Energy Efficient Container Allocation and Migration in Global Cloud Data Centres

<p><strong>Containers have become a popular method for deploying software applications, especially in cloud data centres. Individuals and businesses can rent server space in these centres to run their containerised applications, and even configure them to automatically start or stop in response to fluctuating demand. The task of placing containers onto virtual machines (VMs), and VMs onto physical machines (PMs), is handled by cloud service providers rather than end users. These providers aim to minimise operational costs through efficient resource management. However, because a container’s resource requirements rarely align perfectly with those of a specific VM, careful decisions must be made about placement. Poor allocation can lead to underutilised resources and, as a result, increased energy costs.</strong></p><p>There are a number of existing methods to solve this Container allocation problem. However, they neglected two key aspects of real world data centres. First, many studies do not account for Containers shutting down and leaving the data centre and how this effects future Container allocations. Recent statistics have shown that many containers only run for a few hours at most before completing their task and shutting down. While some existing research considered migrating Containers between different VMs, this is rarely done for energy efficiency reasons. This aspect, combined with frequent container departures, creates an opportunity to migrate remaining containers into the newly freed space. This can help reduce the overall number of virtual VMs and PMs needed at any given time, hereby reducing the overall energy expenditure of cloud data centres.</p><p>A second important but often overlooked aspect is the global distribution of data centres. Reducing the energy consumption of a given data centre is often the immediate approach to reducing the cost of running a set of containers. However power costs, specifically in the wholesale market, are variable and change throughout the day in any particular location and are often lower at night. Containers could be moved between data centres in response to different power prices to achieve a lower overall cost than using a single data centre.</p><p>This thesis presents two novel Genetic Programming Hyper-Heuristic (GPHH) algorithms for energy- and cost-efficient container allocation and migration in cloud data centres. The first algorithm addresses dynamic resource allocation within a single data centre, incorporating real-world considerations such as containers leaving the system and energy-aware migration into newly freed resources. The second algorithm extends the model to a global setting, accounting for geographically distributed data centres with variable, time-dependent power costs. Both algorithms use three co-evolved GPHH trees to guide container-to-VM allocation, VM-to-PM placement, and container migration decisions. To support these innovations, we enhance existing datasets to include container lifetimes and develop a new dataset capturing regional electricity price fluctuations. A custom simulator is also constructed to evaluate dynamic scenarios involving container departures and inter-data-centre migrations. Experimental results demonstrate that our GPHH-based solutions significantly outperform several state-of-the-art methods in reducing energy usage and operational costs, offering practical benefits for real-world cloud infrastructure management.</p>