Microarchitectural Analysis of Pre-Processing Stage in Machine Learning Workloads

As Machine Learning (ML) has become integral for various applications, ML workloads are now important considerations for deployment across diverse use cases, ranging from data centers to edge devices. ML encompasses diverse application fields, including vision, audio, text, and multimodal areas, each involving specific raw data formats that often needs pre-processing to become more interpretable for the models and to ensure a more balanced and standardized data distribution. This stage can also include data augmentation to improve model robustness and performance. Therefore, most ML workloads incorporate a stage, commonly referred to as pre-processing, prior to processing the actual data in complex ML model. As the amount of data size increases at a drastic rate, the preprocessing stage now requires closer attention given its significant computation time. In this paper, we conduct an in-depth microarchitectural analysis of the ML pipeline's pre-processing stage to uncover bottlenecks and utilize a bottom-up approach to deliver valuable insights by identifying code hotspots.