?
Performance Modeling of Data Storage Systems using Diffusion Models
Behavior of various engineering systems exhibits stochastic characteristics. Modeling such systems requires not only estimation of mean values of signals and parameters, but also analysis of their dispersion and higher-order statistics to gain deeper insight into internal system dynamics. Conventional machine learning regression algorithms primarily focus on mean value prediction and often fail to provide the required accuracy in simulation tasks. In contrast, generative artificial intelligence models offer the necessary functionality. In this study, we investigate conditional diffusion models as vendoragnostic surrogates for storage system performance, learning the joint distribution of IOPS and latency conditioned on workload descriptors and system configuration parameters. We utilize a dataset collected from a real data storage system, containing performance observations for cache, HDD, and SSD pools under various data loads and system configurations. Our results indicate that conditional diffusion models provide a robust foundation for data storage digital twins, accurately capturing system behavior and its inherent stochastic nature.