Deep Learning Surrogate Models for Network Simulation

We develop a deep learning surrogate model to accelerate parallel discrete event simulation (PDES) in high-performance computing (HPC) networks, addressing the computational challenges of traditional methods as HPC scales to larger and more complex systems. Our architecture stacks a 1D Convolutional Network, Long Short-Term Memory (LSTM), and a dense layer, trained using synthetic multivariate time series data from the CODES simulation framework. Preliminary results demonstrate the model’s promising performance in predicting application iteration times and its potential for generalization to other HPC workloads. With RSME metrics at an 81.5% improvement and MAE metrics up to 25.6% over baseline statistical methods, our deep learning surrogate approach signifies a shift towards rapid and less resource-intensive predictive models in the HPC simulation landscape.