DEARS: A Deep Learning Based Elastic and Automatic Resource Scheduling Framework for Cloud Applications

摘要

Cloud computing paradigm supports more enterprises to provide satisfactory web services to their clients. However, the bursty and fluctuation of requests challenge the traditional resource scheduling framework. Previous strategies manage the jobs in each virtual machines (VMs) according to the derived historical utilization patterns, where the misalignment on the utilization curves may cause the resource over-prediction and over-provisioning. To better reduce the service latency and the above mentioned problem, we propose DEARS, a Deep learning based Elastic and Automatic Resource Scheduling framework for cloud applications. It gives a proactive and reactive strategy, where the LSTM model is pro-applied to predict the future request demand based on historical workload. The corresponding VM allocation is separately managed by restriction assessment, VM provision, and dynamic consolidation modules. Then the SLAs feedback are iteratively applied to reactively improve the performance of resource allocation. Experiments based on real-life collected data shows the feasibility and efficiency of our framework. The high accuracy of prediction contributes to a more suitable allocation. And a better trade-off between QoS and SLAs in server side is achieved compared with the baselines.