A server consolidationmethod with integrated deep learning predictor in local storage based clouds

摘要

Server consolidation is one of the critical techniques for energy-efficiency in cloud data centers.As it is often assumed that cloud service instances (eg, Amazon EC2 instances) utilize the sharedstorage only. In recent years, however, cloud service providers have been providing local storagefor cloud users, since local storage can offer a better performance with identified price. However,these cloud instances usually contain much more data than shared storage cloud instances.Thus, in such local storage based cloud center, the migration cost can be really high and is in direneed of an efficient resource pre-allocation. If we can predict the resource demand in advance,the migration oscillation will be reduced to minify the migration cost.We have found that thereare some related work about server consolidation based on forecasting. Unfortunately, their latestwork did not consider the background of “local storage” as we mentioned above. At the sametime, some research about local storage did not involve the prediction strategy, which plays asignificant part in server consolidation. To address this issue, this paper proposes Losari, a consolidationmethod,which takes numeric forecasting andlocal storage architecture into consideration.Losari consolidates servers on the basis of the resource demand predicted value using a statisticallearning method.We model the workload from real cloud production environment as a timeseries. Taking deep learning as a frameof reference, multiple deep belief networks integratedwithARIMA model was trained to study the feature of historical workload. The experimental resultshave showed that its average predicted error is only 10.7% in the short term,which is much lowerthan themost commonmodel based on threshold (19.8%) on the same dataset.What ismore, theresults show that Losari not only simulates the true sequences in high accuracy but also scales thecompute resource well, which demonstrated the validity of this integrated deep learning model.