Enhancing SJF and SRTF CPU Scheduling Through Machine Learning-Based Burst Time Prediction: A Comparative Study

Abstract

Accurate prediction of CPU burst time is essential for effective process scheduling, particularly in Shortest Job First (SJF) and Shortest Remaining Time First (SRTF) algorithms. Traditional estimation techniques, such as exponential averaging, can produce substantial prediction errors, resulting in inefficient scheduling, increased waiting and turnaround times, and reduced system throughput. Although machine learning (ML) approaches have recently been investigated for CPU burst time prediction, existing studies are often limited by small datasets, restricted model comparisons, inadequate preprocessing, and insufficiently reproducible implementation frameworks. This study addresses these limitations by developing a comprehensive ML-based framework for CPU burst time prediction using a large-scale real-world workload dataset containing more than 404,000 processes and 29 features. The proposed framework evaluates a broad range of supervised regression algorithms, including Linear Regression, Ridge, Lasso, ElasticNet, Decision Tree, Random Forest, Gradient Boosting, AdaBoost, XGBoost, LightGBM, K-Nearest Neighbors, Support Vector Regression, and Multi-Layer Perceptron. The experimental pipeline incorporates exploratory data analysis, strict train-test separation to prevent data leakage, missing-value imputation, outlier capping, categorical encoding, feature engineering, and standardized performance evaluation using R², adjusted R², RMSE, MAE, and MAPE. Experimental results demonstrate that gradient-boosting approaches substantially outperform conventional and alternative ML models. XGBoost achieved the best performance, attaining a test R² of 0.872, RMSE of 487.3 s, MAE of 196.5 s, and MAPE of 14.2%, compared with approximately 0.70 R² and 30% MAPE for linear baselines. LightGBM achieved a comparable R² of 0.868 while reducing training time by 43%. Feature importance analysis identified AverageCPUTimeUsed, ReqTime, UsedMemory, and NProcs as the dominant predictors, collectively contributing more than 74% of predictive power. The findings indicate strong potential for integrating ML-based burst time prediction into SJF and SRTF scheduling. However, scheduler-level simulations and validation across diverse workload environments are required to quantify actual improvements in waiting time, turnaround time, and throughput.