Exploiting sequence characteristics for long-term workload prediction in cloud data centers

Abstract

In the realm of cloud computing, long sequence prediction of workloads plays a pivotal role, crucial for optimizing resource allocation and enhancing system performance. However, current research of long-sequence workload forecasting faces a series of challenges, mainly due to the high randomness and instability characteristics of long workload sequences, making it difficult for traditional machine learning methods to provide accurate results. Therefore, we designed a novel approach for long sequence forecasting, thoroughly considering the latent characteristics of cloud workload sequences. Initially, we employ convolution kernels of varying sizes to perform multiscale sequence decomposition, better capturing contextual information and periodic features in long sequence. Furthermore, through fast Fourier transformation, we convert one-dimensional sequences into two-dimensional space, leveraging dilated convolutions to extract effective features within the intra-period and inter-period variations. Ultimately, we introduce an attention mechanism, effectively integrating the intra-period and inter-period variation features into the proposed model. Our method has undergone comprehensive evaluation on publicly available datasets from Google, Alibaba, and Microsoft. Experimental results demonstrate superior accuracy and robustness of our model across various workload types, showcasing its excellent adaptability to dynamic and complex workload scenarios.