Research On Method Of Load Forecasting Based On ARAM And Task Scheduling Based On Q-learning

Behind the ability to provide efficient and convenient large data services for edge computing platform,it is an increasing problem of task energy consumption.It is of great significance to accurately predict the task load of edge cloud computing platform(such as how to use CPU and memory)for task scheduling and accurate control of edge computing energy efficiency.Among them,task scheduling is a key point of network technology.Its algorithm performance and network system directly affect the whole network system.Task scheduling algorithm aims to synthesize various resources,including the computing performance of each grid node and the communication performance between nodes,and the communication performance among nodes,so that the total end time of the task is the least.The object of this paper is load forecasting and task scheduling.The main work and innovations of this paper are as follows:(1)Aiming at the load forecasting of edge data center,this paper first introduces the concepts of time series,such as random process,stationarity,autocorrelation function,partial autocorrelation function and white noise,and then introduces the common time series forecasting models: autoregressive model AR(p),moving average model MA(q),autoregressive moving average model ARMA(p,q).The modeling process and principle of ARMA model are introduced in detail:(1)testing and processing of sequence stationarity;(2)model identification and order determination;(3)model parameter estimation and testing;(4)model prediction.In the experimental data set,the t-drive trajectory data set is transformed into the load data set of the mobile data center by using data mining technology,and then aiming at the problem that the algorithm effect of ARMA model with fixed time window step is not ideal,this paper proposes a variable long time window step,and the basic data of each prediction is all the previous historical data,which improves the prediction accuracy.It is compared with the neural network prediction method based on LSTM.(2)In the edge computing task scheduling part,the theory of reinforcement learning is introduced.The Q-learning algorithm based on state matrix and greedy strategy are used to design task scheduling experiment.Firstly,according to the basic principle of reinforcement learning algorithm,the formalized model of edge computing task scheduling is established.Then,based on the representative Q-learning algorithm in reinforcement learning,the algorithm is proposed based on the characteristics of edge computing task scheduling: 1)state space s is based on task matching matrix CVn × m indicates that each state corresponds to a row of a task matching matrix.The row represents the edge cloud task,the column represents the virtual machine,and the update row represents matching the edge cloud task represented by the row to one of the best virtual machines in a column,and the update matrix represents the complete scheduling scheme once.2)Action space a,the number of virtual machines corresponding to each line n,the edge cloud task selects a suitable column,namely virtual machine as the matching object,that is,action.In order to minimize the average completion time,this algorithm matches the edge cloud tasks to VM virtual machine.3)The reward function R,which means(task length / total task length-MIPS of VM /total MIPs)takes a negative number of absolute value,so as to encourage matching with small proportion difference(similarity).4)The design of greedy factor.Compared with the original firstfirst algorithm of cloudsim plus experimental platform,the algorithm proposed in this paper has good performance in average completion time and individual maximum completion time.