Workload, resource and price aware proactive auto-scalar for dynamically-priced virtual machines

Abstract

Proactive Cloud auto-scalers forecast future conditions and initiate scaling response in advance leading to better service quality and cost savings. Their effectiveness depends on the forecast accuracy and penalty due to miss prediction. However, such solutions assume fixed prices for virtualized Cloud resources to be provisioned. Hence, they are unable to benefit from dynamically-priced resources such as Amazon Spot Instances which are introduced by Cloud providers to deal with fluctuating workloads cost effectively. Moreover, users have the risk of losing resources when the dynamically-adjusted market price of resources exceeds the user-defined maximum bid price. Therefore, proactive auto-scalers should also forecast market price of dynamically-priced resources to minimize the cost further while retraining service quality. However, predicting the market price (to set the maximum bid price) is quite complicated given highly varying workload and resource demands. We present a proactive auto-scalar for dynamically-priced virtual machines by combing the workload and resource prediction capabilities of an existing auto-scalar named InteliScaler, and a novel technique for forecasting Spot price. We retrieve Spot price history from Amazon and use it to forecast the future prices using Recurrent Neural Networks. Next, we selected the maximum price for a given decision window as the bid value to make Spot request. To demonstrate the utility of the proposed solution, we tested the performance of the enhanced auto-scaler using a synthetic workload generated using the Rain toolkit and the RUBiS auction site prototype. Proposed auto-scaler with dynamically-priced virtual machines reduced the total cost by ~75% compared the same auto-scalar with fixed priced instances. Moreover, no noticeable change in service quality was observed