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

TL;DR

This paper introduces MambaCPU, a novel neural network model utilizing state space models and attention mechanisms to improve CPU performance prediction accuracy, addressing data diversity and correlation challenges.

Contribution

The paper presents MambaCPU, a new model leveraging global dependency exploration and attention mechanisms, along with a new dataset PerfCastDB for enhanced CPU performance forecasting.

Findings

MambaCPU outperforms existing prediction methods on PerfCastDB.

The model effectively captures complex multivariate correlations.

Open-sourced dataset and code facilitate further research.

Abstract

Forecasting CPU performance, which involves estimating performance scores based on hardware characteristics during operation, is crucial for computational system design and resource management. This research field currently faces two primary challenges. First, the diversity of CPU products and the specialized nature of hardware characteristics make real-world data collection difficult. Second, existing approaches, whether reliant on hardware simulation models or machine learning, suffer from significant drawbacks, such as lengthy simulation cycles, low prediction accuracy, and neglect of characteristic correlations. To address these issues, we first gathered, preprocessed, and standardized historical data from the 4th Generation Intel Xeon Scalable Processors across various benchmark suites to create a new dataset named PerfCastDB. Subsequently, we developed a novel network, MambaCPU (MaC), as the baseline model for the PerfCastDB dataset. This model employs the mamba structure to explore global dependencies and correlations among multiple characteristics. The use of intra- and inter-group attention mechanisms further refines correlations within and between characteristic groups. These techniques enhance MaC's capability to analyze and mine complex multivariate correlations. Comparative experiments on the PerfCastDB dataset demonstrate that MaC surpasses existing methods, confirming its effectiveness. Additionally, we have open-sourced part of the dataset and the MaC code at \url{https://github.com/xiaoman-liu/MaC} to facilitate further research.