Energy-Saving Of Deep Learning Tasks Based On CNTK

In recent years,deep learning has been successfully applied to various issues.The success of deep learning is due to the increasing amount of data accumulated,and the multi-layer artificial neural layer has high characterization capabilities for data,GPUs also play a crucial role in deep learning by significantly reducing the time spent of training model parameters.In recent years,various GPU clusters for deep learning have been established in various companies and scientific research institutions.GPU clusters typically consist of hundreds or thousands of nodes in large enterprises.Such a huge cluster runs for a long time,It will consume a lot of energy,increase the instability of the system and the operating costs of the company.Therefore,reducing energy consumption in large-scale GPU clusters can not only economically reduce expenditures for companies and research institutions,but also achieve true environmental protection.GPU or general computer clusters do not consider the issue of saving energy when running.When the cluster is idle for a period of time,all nodes will be in a long empty load state,or the utilization rate of precious resources such as GPU of many nodes is relatively low,so the energy efficiency is relatively low.The research on cluster multi-GPU training models or predictive models is basically performance-based and based on the special nature of the deep learning task itself.Therefore,most of the scheduling algorithms applied directly to GPU clusters cannot achieve good results,cause a lot of energy waste.This thesis focuses on the above issues,Based on CNTK,the deep learning related energy-saving scheduling technology of GPU clusters and containerized resource scheduling is the main research object,and in-depth analysis of the insufficiency of existing GPU clusters and scheduling methods.A solution to reduce energy consumption is proposed for GPU cluster training and containerized deep learning task scheduling.The main work of this thesis is as follows:(1)First,this thesis analyzes the feasibility and solution of energy-saving in deep learning GPU clusters,defines the energy consumption measurement model.and improves a dynamic energy-saving scheduling algorithm for deep learning training GPU clusters.(2)Secondly,this thesis proposes a new integration approach for deep learning container resource minimum energy consumption and a algorithm of DEDDEP for newly emerged deep learning scenarios based on containerized resource scheduling.(3)Finally,Tested a lot of deep learning data by designing test cases.After a large number of real data tests,the proposed method has been proved to be effective on GPU clusters,and for deep learning containerized resource scheduling scenarios,the energy-saving effect of source container selection based on dual thresholds of GPU and memory usage and DGDDEP algorithm in container allocation are verified.