SD-MoE: Scenario-Driven MoE Forecasting for Intelligent Elastic Scaling in Cloud Clusters
Year: 2026
Type: conference-paper
Abstract: Traditional CPU capacity planning in cloud computing often depends on static heuristic rules, which struggle to accommodate temporal dynamics and heterogeneous workload patterns, frequently resulting in resource overprovisioning or performance degradation during demand surges. To address these challenges, we propose SD-MoE, a learning-based framework that integrates scenario-aware sparse expert mechanisms into CPU usage forecasting. The framework comprises four key components: (1) a dual-branch input pipeline that jointly models historical periodicity and real-time telemetry to capture long-term trends and short-term fluctuations; (2) a cross-dimensional transposed embedding module that preserves structured temporal-metric dependencies while mitigating noise from heterogeneous scales; (3) a hierarchical dynamic-routing Mixture of Experts (MoE) that replaces the conventional feed-forward network with four timestamp-activated expert types (weekday-daytime, weekdaynighttime, holiday-daytime, holiday-nighttime), enabling timespecific behavior modeling while retaining global knowledge; and (4) expert grouping and semantic constraints aligned with business scenarios to improve interpretability and facilitate operational integration. Experiments on large-scale production traces demonstrate that SD-MoE significantly enhances peak-load sensitivity and reduces overprovisioning. Notably, it achieves an average R-value of 0.3845, outperforming the strongest baseline by over 13%.
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7171984108