Abstract
The adoption of cloud services continues to grow, with applications ranging from simple web tools to complex content management systems increasingly relying on virtual compute instances powered by Central Processing Units (CPUs). The emergence of generative AI has further amplified the demand for high-performance computing resources such as Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs). However, current cloud billing models typically follow a “pay-as-you-order” paradigm rather than a true “pay-as-you-use” model, often leading to inefficient resource utilisation and idle time between operations. In this paper, we present a smart resource distribution method for CPUs, which is also extendable to GPUs. Our method supports containerised applications with diverse computational requirements, from those that need multiple CPUs to those that can run efficiently on a fraction of a single CPU. We evaluated the approach using a simulated data pipeline application without disk I/O, comparing it against both static allocation and the default Completely Fair Scheduler (CFS). The results demonstrate improved resource utilisation and overall performance of the application. Future work includes a comparison with machine learning–based scheduling techniques to assess potential trade-offs in computational overhead.