Few-shot HPC application runtime prediction

Si Chen, Simon Garcia De Gonzalo, Avani Wildani

Abstract: Precise runtime prediction for HPC jobs is essential for streamlined hardware/software co-design, resource allocation, and assessing the impact of hardware alterations. Existing runtime prediction methods, however, are generally application and architecture-specific, hindering their broad applicability. In response, we propose a novel meta-learning and simulation-based model that accommodates a wide range of applications and architectures. This method efficiently addresses new runtime challenges using only a limited number of samples. As demonstrated by our experiments, with just ten training samples, this model attains an average MAPE of 19% on the SPEC CPU 2006 benchmarks.