Workload Prediction of Cloud Workflow Based on Graph Neural Network

With the continuous expansion of cloud computing market, the problem of low utilization rate of cloud computing resource has become increasingly prominent, because cloud computing vendors can not schedule a large number of server cluster effectively as before. Improving the utilization rate of cloud resources can not only improve the net profit of cloud computing manufacturers, but also reduce the time cost and economic cost of cloud computing users. In addition to resource scheduling, the current research on cloud workflow load is still focused on single task or single instance prediction, and even the data sets used are simulation data. This paper aims to predict workload of cloud workflow resources to make the cloud computing resources get better scheduling, and ultimately facilitate all relevant personnel in the cloud computing market. Firstly, compared with task and single instance, cloud workflow can get more context information. Secondly, in order to make this research more practical, this paper selects Alibaba cluster data V2018 released by Alibaba in 2018 as our research object. Thirdly, based on the graph structure characteristics of cloud computing workflow, this paper selects the Graph Neural Network (GNN) architecture which closely fits the graph structure to predict the load of cloud computing workflow, and specifically selects the homogeneous Graph Convolution Neural Network and Graph Attention Neural Network and heterogeneous GCN as our prediction algorithm. And it describes how cloud workflow is modeled as homogeneous graph and heterogeneous graph in detail. Finally, the algorithm in GNN is used to classify and predict Ali data with workflow length ranges from 4 to 12 separately and combined, and predicts the last and penultimate tasks of each length workflow. Besides, all the data from 4 to 12 are combined into one data to predict the last and penultimate tasks.