SWIFT framework cuts cloud workload errors by 31% and latency 80%

Accurate cloud workload forecasting is critical for efficient resource management but remains difficult due to volatile traffic and sudden bursts. Traditional wavelet-based methods preserve temporal locality but rely on rigid fixed bases that struggle with complex patterns, and they often ignore spatial dependencies between variables. To address this, researchers from Shanghai Jiao Tong University propose SWIFT (Spatio-temporal Wavelet Integrated Forecasting Framework), a pure convolutional architecture designed for high-efficiency workload prediction.

SWIFT introduces two key components: a Learnable Cascaded Wavelet Path that transforms fixed wavelet bases into adaptive convolutional operators for data-driven feature extraction, and a Multivariate Interaction Module that sequentially models inter-variable spatial and intra-variable feature interactions to stabilize noisy states. In extensive experiments, SWIFT achieves state-of-the-art accuracy with linear O(L) complexity, reducing prediction error by up to 31.04% while cutting inference latency by 79.74%. This makes it particularly suitable for real-time cloud resource allocation, potentially saving costs and improving performance under varying loads.