Efficient file prefetching is essential for optimizing input-output performance in modern operating systems. The readahead mechanism in Linux plays a crucial role in this process by predicting future file access patterns and pre-loading data. However, existing heuristics often struggle to adapt dynamically to varying workloads or other machine learning based approaches might struggle with system-level tasks like readahead because it does not capture temporal features effectively. Current readahead algorithms lack the ability to predict real-time changes in file access patterns, leading to inefficiencies in file system performance, especially under diverse workloads. In this approach, temporal and non-temporal features are utilized, applying models such as Extreme Gradient Boosting and Long Short-Term Memory networks for readahead size prediction. This approach compared the performance of Long Short-Term Memory and Extreme Gradient Boosting for system-level predictive tasks. Long Short-Term Memory outperformed Extreme Gradient Boosting in sequential data handling, achieving higher accuracy but at the cost of greater computational overhead and latency. Extreme Gradient Boosting, though faster and more resource-efficient, struggled with capturing long-term dependencies, still giving an accuracy similar to Long Short-Term Memory. This approach demonstrates the potential for machine learning models to enhance file prefetching mechanisms in Linux, offering a more adaptive and efficient solution for handling diverse file access patterns. Long Short-Term Memory and Extreme Gradient Boosting models have demonstrated impressive performance metrics, with Long Short-Term Memory achieving an \(\textrm{R}^{2}\) value of 0.9886 and Extreme Gradient Boosting reaching 0.9785. These results indicate that both models significantly outperform Kernel Machine Learning’s accuracy of 95.5%.