Modelling cloud service latency and availability using a deep learning strategy

Low latency and high availability in cloud services give users satisfactory response time and guarantee stability to request they make to services that are hosted in the cloud, thus increasing the usability and reliability of cloud services. On the other hand, high latencies and poor availability will cost businesses their customers due customers dissatisfaction, thus losing customers to competitors. This situation noticeable in e-commerce businesses where real-time response and decision making within seconds are critical for business service delivery. Therefore, latency and availability are important parameters in the Service Level Agreement for cloud users when choosing Cloud Services Providers (CSPs). But the challenge for businesses is that they do not have their in-house mechanism that can accurately predict the required latency and availability for their requirements. Companies only rely on the CSPs tools for estimating resource requirements, which is biased towards the CSPs business model. In this paper, we developed a deep learning algorithm that predicts the latency and availability of cloud services using real-time live data from three CSPs. We designed and implemented experiments on Amazon Web Services, Alibaba Cloud and Tencent Cloud in Beijing University of Posts and Telecommunications to run compute instances across the United States, Europe and Asia Pacific regions. In each cloud platform, five servers were used that resulted in 30,815,100 invocations of http and ping operations for 6 weeks. The algorithm used the data on hourly, daily and weekly basis as historical network data to predict latency and availability. We used MATLABs deep learning toolbox for the implementation of our algorithm and the results showed that the prediction is usually above 90% accurate as compared with the data obtained. The results also revealed that latency performance depends on the locations of users and the availability depends on number of availability zones used.