Exploring Machine Learning Models with Spatial-Temporal Information for Interconnect Network Traffic Forecasting

Abstract

Interconnect networks are an essential component of high-performance computing (HPC) systems. To study large-scale networking systems, parallel discrete event simulation (PDES) has been widely used to simulate real-world HPC behaviors. However, PDES simulation requirements and computational complexity are increasing rapidly, making it challenging to achieve accurate results. Therefore, researchers have been exploring a surrogate-ready PDES framework that utilizes machine learning-based surrogate models to accelerate PDES. In this paper, we present our vision and initial step to leverage machine learning models to utilize spatial-temporal information to forecast interconnect network traffic. The preliminary results show that it is promising to explore machine learning models for interconnect network traffic forecasting.