A Recurrent Neural Network Based Approach for Web Service QoS Prediction

QoS (Quality of Service) is a series of widely used indicators for networked user-oriented online services such as SOAP/JSON based web services, serverless microservices or function services in the modern cloud market. Of which, response time and throughput are most important measurements which almost determine the whole user experience caused by underlying factors including network delay or congestion, fluctuating service computation performance at different system workload level, or constrained by cascading resource or service dependencies. It's valuable to accurately predict QoS indicators under complex scenarios and conditions combination from archived history service experience dataset. In this paper, two subsets of real-world dataset named WS-DREAM were used, which collected 339 users' records on 5825 web services. Based on related work and descriptive analysis, two QoS (response time and throughput) prediction models were proposed for classification and regression problems. A custom recurrent neural network(RNN) was proposed, which comprises stacked multiple LSTM(Long term and Short Term Memory) layers, several regularization techniques were used. Extensive experiments were conducted to verify the accuracy and efficiency of deep learning based approach comparing with 6 competitive machine learning approaches, including decision tree, AdaBoost, multilayer perceptron, XGBoost, LightGBM, and CatBoost. Results showed the prediction accuracy of most algorithms could reach more than 90%, and the proposed RNN based network showed a superior performance both at accuracy and efficiency than other approaches.