Distributed deep neural network performance modeling method based on an instruction queue

Abstract

The invention discloses a distributed deep neural network performance modeling method based on an instruction queue. The method is characterized by comprising hardware performance parameters; A neuralnetwork structure, a bottom layer computing architecture, a data transmission protocol, link bandwidth characteristics and a server GPU topological structure are used for carrying out layered mappingsplitting and segmented computing on a deep neural network, then an instruction queue is used for estimating training iteration time consumption, and meanwhile the data interaction condition betweenhardware is output. According to the distributed deep neural network performance modeling method based on the instruction queue, software and hardware characteristics are considered at the same time;Time consumption analysis is carried out by using an instruction-level queue model, so that one-time iteration time consumption estimation of deep neural network training and analysis of each hardwaredata interaction process are realized, and the method is suitable for different hardware environments (different servers, different types of GPUs and different block numbers of GPUs) and different neural networks.

Description

technical field [0001] The invention belongs to the technical field of performance modeling based on a specific calculation model, and in particular relates to a modeling method for the performance of a deep neural network trained on a single or multiple graphics processing units (GPUs). Background technique [0002] The central processing unit (Central Processing Unit, CPU) is the computing core and control core of a computer, and the graphics processing unit (Graphics Processing Unit, GPU) is a microprocessor for image computing work. Compared with CPU, GPU has more computing units. Since the development of GPU general-purpose computing technology, GPU has been widely used in large-scale computing tasks, especially in the field of deep learning. [0003] Deep Neural Network (DNN) refers to an artificial neural network (Artificial Neutral Network, ANN) with multiple hidden layers. Its concept was proposed by Jeffrey Hinton's research group at the University of Toronto in 20...

AI Technical Summary

Problems solved by technology

At present, the performance modeling method of deep neural network trained on GPU is mainly carried out by collecting the underlying data through experiments: for example, in the article "Performance modeling and evaluation of distributed deep learning framework on gpus[C]" included in IEEE in 2018, for Three different deep learning frameworks have constructed a time-delay model of one iteration of the convolutional neural network, but the model construction depends on the experimental results, and it cannot give an estimate of the time-consuming of one iteration of neural network training; there are also theoretical calculations The practice of estimating performance: For example, the article "Paleo: A Performance Model for Deep Neural Networks [C]" included in the conference ICLR (In Proceedings of the International Conference on Learning Representations) in 2017 was constructed for different networks and distributed hardware environments Deep neural network performance model, which is mapped to specific software, hardware, and communication strategy spaces, is used to explore the scalability of deep learning systems, but its error can only be controlled within 30%

Images