Research On Random Forests Online Learning For Parallel Rendering Load Balancing

The rapid development of modern graphic applications poses more and more severe challenges to the super-large-scale complex scenes rendering,high-resolution display,high-realistic user experience,and real-time rendering efficiency.Although the performance of computer hardware has been greatly promoted in recent years,it is still hard to satisfied the increasing requirement.As a specific application of parallel computing in the field of graphics,the parallel graphics rendering system has become one of the effective solutions to resolve the above problems.However,the assignment of drawing tasks among multiple computing nodes has always been a bottleneck restricting the overall performance of the parallel rendering system.Therefore,research on how to realize rendering load balance has become a key factor affecting the finally rendering effect.This paper takes the rendering time of rendering nodes as the load measurement,and regards the accurate prediction of rendering load as the key issue of load balance.By predicting the load in advance,the rendering tasks can be assigned reasonably to achieve a load balance effect.In this paper,Random Forests(RFs)are proposed to express the high dimensional nonlinear relationship between rendering features and rendering time.Aiming at the problem of continuous generation of rendering data and the target concept drifts,an adaptive load balancing framework of “Prediction+Correction+Learning” is proposed.The online learning mechanism of RFs in rendering data stream is focused on,which be used to improve the regression prediction performance of the model and enhance the model's adaptability in the dynamic data stream.The specific work and innovation of this paper are as follows:(1)To enhance the accuracy of regression prediction in rendering data stream,an online learning approach for RFs based on leaf node weight is proposed,which is called OWL-RFR(Online Weight Learning for RFs Regression).The basic idea is to make the leaf nodes obtain a long-term dependency based on the correlation of the sequence data in the stream without changing the structure of off-line trained RFs,and optimize the weights of leaf nodes through gradient descent.Thus,an effective long-term memory is endowed to RFs,which makes the model can continuously learn during the prediction process of the data stream,and the more and more accurate results are obtained.The experimental results show that the OWL-RFR method can better fit the continuous rendering data stream,improve the accuracy of regression prediction,and has better convergence and stability.(2)To improve the adaptability of the model in the dynamic rendering data stream,a long and short-term memory adaptive online RFs method called ALSM-RFR(Adaptive Long and Short-term Memory online RFs Regression)is proposed.Due to the occurrence of the target concept drift caused by the changes of the scene and the user interaction in the parallel rendering system,the OWL-RFR method may not be able to adapt to the constantly changing dynamic rendering data stream.Hence,an adaptive memory activation mechanism is designed to enable RFs to adaptively switch the memory used for prediction between different memory modes when predicting stationary or non-stationary data flow.The ALSM-RFR method uses both leaf-level weight and tree-level weight to continuously accumulate knowledge of different periods during the prediction process and endows the model with long-term memory and short-term memory at the same time,indeed,a hybrid memory is synthesized based on them.The experimental results show that compared with the OWL-RFR,the ALSM-RFR obtains a better prediction effect in concept drift-oriented rendering data stream prediction.In the meantime,the ALSM-RFR has good convergence and stability,and the optimization strategy of batch gradient and stochastic gradient descent also effectively reduces the impact of hyper-parameters.(3)To enhance the model's adaptability in the complex drift rendering data stream,an online local reconstruction RFs method called ORB-RRF(Online Rebuilding Regression Random Forests)is proposed.Although the usage of long and short-term memory has improved the adaptability of RFs to a certain extent,the RFs that keep the offline training structure unchanged are difficult to adapt to drastic changes when face with more complex dynamic and changing data stream.Therefore,the ORB-RRF method is proposed to adapt to complex dynamic continuous data streams with multiple types of concept drift.On the one side,by comparing the performance of leaf nodes and their father nodes online,the leaf nodes are pruned online to recorrect performance degradation caused by the improper division of the feature space by the leaf nodes and suppress the model overfitting;On the other side,the leaf nodes collect these samples with poor prediction effects dynamically during prediction process,and the changes between features space contained in leaf node after it adds new samples and its original feature space are compared layer by layer from bottom up.Through the backtracking operation,the ancestor node that can cover the new features space is located and re-split from this node to obtain a new subtree to replace the original branch.The experimental results show that in the data stream prediction for complex concept drifts,the ORB-RRF method has a significant performance improvement over OWL-RFR and ALSM-RFR methods.Through the online rebuilding of the structure,the RFs have the adaptive ability of continuous optimization.