基于梯度提升决策树的材料计算时间预测模型

Material Calculation Time Prediction Model Based on Gradient Boosting Decision Trees

材料计算运行时间预测对提升作业调度效率和新材料研发有着至关重要的作用,传统集群作业运行时间预测模型的精度较差,在领域的可用性较低。为此,提出一种基于梯度提升决策树的作业预测模型,结合领域知识和相关文献对VASP作业日志数据进行清洗,对选择的特征进行重要性评估,然后在不同数据量、不同样本分布数据条件下进行实验,将该模型与使用传统机器学习方法的模型进行比较。实验表明,所提方法的平均绝对百分比误差在不同条件下均低于传统机器学习方法,且综合作业运行时间的预测误差为4.28%,优于RunningNet方法的10.3%,证明了所提模型对材料计算运行时间预测精度更高,对提升作业调度效率和加快新材料研发作用更大。

Prediction of material calculation run time plays a crucial role in improving job scheduling efficiency and new material research and development.Traditional cluster job run time prediction models have poor accuracy and low availability in the field.To this end,a job predic⁃tion model based on gradient boosting decision tree is proposed,which combines domain knowledge and relevant literature to clean VASP job log data,evaluates the importance of selected features,and then conducts experiments under different data sizes and sample distribution con⁃ditions.The model is compared with models using traditional machine learning methods.The experiment shows that the average absolute per⁃centage error of the proposed method is lower than that of traditional machine learning methods under different conditions,and the prediction error of comprehensive job running time is 4.28%,which is better than the RunningNet method's 10.3%.This proves that the proposed model has higher accuracy in predicting material calculation running time,and has a better effect on improving job scheduling efficiency and acceler⁃ating new material research and development.