Research On Artificial Intelligence For IT Operations In Dynamic Edge Networks

The computing and storage capability of edge network provides an efficient solution for delay sensitive and context sensitive application services.Due to the instability of wireless channel transmission,the heterogeneity and mobility of computing nodes,and the randomness of task request arrival,the edge network shows obvious dynamic characteristics,which bring many difficulties and challenges to the operation of the edge network.This paper focuses on two intelligent operation strategies in the dynamic edge network,including the parallel task scheduling and multi-time series anomaly detection.The dynamic characteristics of edge networks greatly limit the performance of most kinds of heuristic algorithms,which need to model the problem environment accurately within the algorithm.We proposed a Multi-task Deep reinforcement learning Task Scheduling algorithm(MDTS).We mapped the parallel task scheduling problem in the dynamic edge network to the multi-task deep reinforcement learning problem,where each output branch of the multi-task learning corresponds to the scheduling scheme of each child task.MDTS does not require much prior knowledge of the environment or computing nodes,but directly learns the end-to-end allocation scheme,and understands the resource competition and cooperation relationship among parallel tasks through the shared neural network in the model.In addition,we also designed an appropriate reward function to optimize multiple performance metrics simultaneously,supporting the customization requirements of each metric in different scenarios.Adequate experiments showed that compared with the classical least connection scheduling algorithm and the heuristic particle swarm optimization algorithm,MDTS significantly improved the total reward value of task scheduling.In the actual edge network operation,the performance of task scheduling influenced not only by the design of scheduling algorithm,but also by the selection of candidate computing nodes in good health.Multi-time-series anomaly detection algorithms can be used to solve this problem.We proposed a Multi-Time-Series Anomaly Detection Algorithm Based on Graph Neural Network and Variational Autoencoder(Graph-VAE)in this paper.Different from other detection algorithms,Graph-VAE explicitly models the correlation between various KPIs(Key Performance Indicators)through the graph-attention layer,and provides Pearson correlation coefficients between each KPI curve as a priori knowledge.In addition,Graph-VAE also realized the automatic selection of abnormal threshold through POT(Peaks-Over-Threshold)algorithm.The experiments proved that the F1-score of Graph-VAE reached 0.9403 after automatically selecting abnormal threshold,which beat other advanced multi-time-series anomaly detection algorithms.At the same time,we have also proved the effectiveness of the Graph-VAE algorithm in abnormal diagnosis.