Toward Scalable Empirical Dynamic Modeling

Empirical Dynamic Modeling (EDM) is an emerging non-linear time series analysis framework that allows prediction and analysis of non-linear dynamical systems. Although EDM is increasingly adopted in various research fields, its application to large-scale data has been limited due its high computational cost. This article describes our ongoing efforts toward accelerating EDM computation using HPC technologies such as GPU offloading and parallel processing using. We describe mpEDM, a massively parallel implementation of EDM designed for GPU-accelerated supercomputers, and kEDM, a performance-portable implementation of EDM based on the Kokkos performance portability framework. Furthermore, we present our ongoing work toward porting EDM to NEC’s Vector Engine processor and carry out a preliminary performance evaluation.