MetaCast: Generalizing HPC Application Runtime Prediction

High-Performance Computing (HPC) runtime prediction is essential for optimizing hardware/software co-design, allocating resources, and evaluating hardware modifications. Traditional runtime prediction methods are often constrained to specific applications and architectures. To address these limitations, we introduce MetaCast, a novel approach that combines meta-learning techniques with architecture simulation for HPC runtime prediction. With an extensive spectrum of HPC applications across diverse architectural platforms, MetaCast efficiently tackles new runtime prediction challenges using only a limited number of samples for training. As demonstrated by experiments, with just ten training samples, MetaCast attains an average Mean Absolute Percentage Error (MAPE) of just 18% on the SPEC CPU 2017 benchmark and 25% on real applications.