MambaCPU: Enhanced Correlation Mining with State Space Models for CPU Performance Prediction

Abstract

CPU performance prediction, which involves forecasting the performance scores of a CPU based on its hardware characteristics during the operation process, is a critical technology for computational system design and resource management. However, this research field currently faces two significant challenges. First, collecting real-world data is challenging due to the wide variety of CPU products on the market and the highly specialized nature of relevant hardware characteristics. Second, existing methods based on hardware simulation models or machine learning exhibit notable shortcomings, such as lengthy simulation test cycles, low prediction accuracy, and the ignoration of characteristic correlations. To bridge these gaps, we first collect, preprocess, and standardize historical data from the 4th Generation Intel Xeon Scalable Processors across multiple benchmark suites to create a new dataset, named PerfCastDB. Subsequently, we design a novel network MambaCPU (MaC) as the baseline for the PerfCastDB dataset. This model leverages the mamba structure to mine the global dependencies and correlations between multiple characteristics. The intra- and inter-group attention mechanisms are subsequently utilized to refine the correlations within and across the characteristic type groups. These techniques enhance the analysis and mining capability of MaC for the complex multivariate correlations. Comparative experiments on the PerfCastDB dataset demonstrate that MaC achieves superior results compared to existing methods, validating its effectiveness. Furthermore, we have open-sourced a portion of the dataset and the MaC code at https://github.com/xiaoman-liu/MaC to facilitate the subsequent research.