FP-MLP: A Frequency Domain Patch-Based MLP Model for GPU-Dominated Cloud Workload Prediction

Cloud Data Centers are typically equipped with various types and performance levels of GPUs, making it crucial to effectively leverage these heterogeneous resources when deploying deep learning tasks. However, differences in GPU performance pose complex resource management challenges. Accurate workload forecasting can help address these challenges. Many existing prediction models employ Multi-Layer Perceptrons (MLPs) due to their simplicity, computational efficiency, and general applicability. Nevertheless, most MLP-based methods tend to prioritize low-frequency features, neglecting high-frequency features that are essential for capturing fine-grained patterns in resource utilization data. Our analysis reveals that this bias primarily stems from the dominance of low-frequency features in the dataset, causing the model's attention to disproportionately focus on them. To tackle this issue, we propose a frequency domain patch-based MLP model called FP-MLP, which partitions the frequency series into small patches, thereby balancing the representation of high and low-frequency features across different bands. This approach enables the model to fully capture hidden features in high-frequency data, thus improving predictive accuracy. Extensive experiments demonstrate that the FP-MLP model consistently outperforms various state-of-the-art baseline models with stable performance. For instance, on the Alibaba dataset, compared to the best baseline model, the FP-MLP model reduces the Mean Squared Error (MSE) by 8% and the Mean Absolute Error (MAE) by 17%; meanwhile, it also achieves competitive results on the Google.