多参数融合的深度学习算子及模型性能预测方法OA

Efficient resource allocation is a critical factor for enhancing the resource utilization of deep learning models in cloud computing data centers.Static resource allocation methods,which allocate re-sources by analyzing the computational demands of models,have emerged as one of the effective approa-ches to improve resource utilization.However,they still suffer from issues of insufficient efficiency and flexibility.To address this,we propose a multi-parameter fusion-based performance prediction method for deep learning operators and models.Firstly,an automated operator performance testing method is proposed based on operator parameters.Subsequently,by integrating data dependencies among opera-tors and hardware performance parameters,the method enables inference performance prediction for convolutional neural networks(CNN)and recurrent neural networks(RNN).Experimental results demonstrate that,on a single platform(either a single CPU or GPU),the proposed method achieves an average prediction error of 6.2%for CNN and RNN inference performance,representing a 2.3 percen-tage point reduction compared to SLAPP.On heterogeneous platforms(CPU+GPU),the maximum error does not exceed 10%.The proposed method enhances the accuracy of inference performance pre-diction,providing valuable insights for resource allocation in cloud computing data centers.