摘要
将负载预测问题抽象为机器学习中的回归问题,同时考虑到复杂场景中的绘制参数较多所带来的特征维度高、各特征之间可能存在较强的相关性,以及实时应用对效率的要求等,在繁多的回归算法中,选择XGBoost作为学习和预测负载的机器学习模型,实现上述负载平衡划分的关键在于利用建立好的机器学习模型对绘制帧在特定的子屏幕下的负载进行准确且高效地预测。以并行绘制系统中的四分屏作为研究对象,以测试集上的均方误差(MSE)作为预测准确率的评价标准,基于帧间相关性的方法在测试集上的MSE为3.56,XGBoost算法在测试集上的MSE为2.06,其预测更为准确。
Parallel rendering technology is a common solution for rendering complex scenes in real time.The problem of load balancing is the important problem in parallel rendering system research.Some real-time applications based on interactions require a high frame rate for the rendering system to output high quality images.In this case,load balancing is very important.At present,there are many solutions to the load balancing problem in parallel rendering systems.But they have some disadvantages in real-time and stability,especially in the case of frequent user interactions,it is impossible to fundamentally achieve high-reality real-time rendering of large-scale complex scenes.In recent years,machine learning has been widely used in computer graphics,more and more researchers have begun to use machine learning methods to solve regression problems.The overhead prediction problem is abstracted into the regression problem in machine learning.At the same time,considering the high rendering parameters in the complex scene,the feature dimension is high,there may be strong correlation between the features,and the requirements of efficiency in real-time application,etc.In a variety of regression algorithms,XGBoost is selected as the machine learning model for learning and predicting overhead.The key to achieve the above load balancing division is to use the established machine learning model to accurately and efficiently predict the overhead of rendering frame under a specific sub-screen.The four-tile screen in the parallel rendering system is used as the research object,and the Mean-Square Error (MSE) on the test set is used as the evaluation standard of the prediction accuracy.The MSE of the method based on the inter-frame correlation is used on the test set is 3.56,and the MSE of the XGBoost algorithm on the test set is 2.06,the latter is more accurate.
作者
贾文娟
JIA Wen-juan(College of Computer Science,Sichuan University,Chengdu 610065;Sichuan Dazhisheng Software Company,Chengdu 610065)
出处
《现代计算机》
2019年第1期63-66,86,共5页
Modern Computer
基金
国家重大仪器开发专项(No.2013YQ490879)
