Files
Abstract
Resource provisioning is essential to optimize cloud operating costs and the performance of cloud applications. Understanding job arrival rates is critical for predicting future workloads to determine the proper amount of resources for provisioning. However, due to the dynamic patterns of cloud workloads, developing a model to accurately forecast job arrival rates is a challenging task. Previously, various prediction models, including Long-Short-Term-Memory (LSTM), have been employed to address the cloud workload prediction problem. Unfortunately, the current state-of-the-art LSTM model leverages recurrences to make a prediction, resulting in increased complexity and degraded computational efficiency as input sequences grow longer. To achieve both higher prediction accuracy and better computational efficiency, this work presents a novel time-series forecasting model for cloud resource provisioning, called WGAN-gp (Wasserstein Generative Adversarial Network with gradient penalty) Transformer. WGAN-gp Transformer is inspired by Transformer network and improved WGAN (Wasserstein Generative Adversarial Networks). Our proposed method adopts a Transformer network as the generator and a multi-layer perceptron network as a critic to improve the overall forecasting performance. WGAN-gp also employs MADGRAD (Momentumized, Adaptive, Dual Averaged Gradient Method for Stochastic Optimization) as the model’s optimizer due its ability to converge faster and generalize better. Extensive experiments on the various real-world cloud workload datasets show improved performance and efficiency of our method. In particular, WGAN-gp Transformer shows 5× faster inference time with up to 5.1% higher prediction accuracy than the state-of-the-art workload prediction technique. Such faster inference time and higher prediction accuracy can be effectively used by cloud resource provisioning and autoscaling mechanisms. We then apply our model to cloud autoscaling and evaluate it on Google Cloud Platform with Facebook and Google cluster traces. We discuss the evaluation results showcasing that WGAN-gp Transformer-based autoscaling mechanism outperforms autoscaling with LSTM by reducing virtual machine over-provisioning.