Mixed contrastive transfer learning for few-shot workload prediction in the cloud

Accurate prediction of workloads is the key to influencing elastic resource management scaling in cloud platforms. There are a large number of workloads with few sampling points and short-run cycles in cloud platforms. They have the problem of insufficient training data for deep learning, known as Few-Shot Learning. Few-shot learning poses challenges for accurate workload prediction. To address this problem, this paper proposes a mixed contrastive transfer learning for few-shot workload prediction in the cloud. The method utilizes generated samples and source domain data to enhance the information content of the few-shot sequences. The issue of low prediction accuracy resulting from insufficient a prior information of limited workload sequences is solved. First, in the data mixup stage, the mixup method is used to mix different samples to create new samples. Then, in the contrastive representation transfer stage, augmented samples are chosen to create positive sample pairs with the mixed samples. The representation capability of the target model is enhanced by aligning the representational relationships between target and source models described by contrastive learning. Finally, we conducted extensive experiments on Google cluster trace and Alibaba cluster-trace-v2018 to demonstrate the model's generalizability. The experimental results show that the proposed method not only generates samples with similar distributions but also significantly improves the feature representation ability of the few-shot prediction model. Compared with the SOTA prediction method, the Mean Absolute Error and Mean Square Error of the proposed method are reduced by 20.1% and 34.2%.