Korali: a High-Performance Computing Framework for Stochastic Optimization and Bayesian Uncertainty Quantification
share
We present a modular, open-source, high-performance computing framework for data-driven Bayesian uncertainty quantification and stochastic optimization. The proposed framework (Korali) is well suited for the non-intrusive sampling of computationally demanding engineering and scientific models. The framework's distributed-execution engine allows for the efficient execution of massively-parallel computational models while providing fault tolerance and load balancing mechanisms. In this paper, we present our framework's design principles and explain its flexibility in allowing scientists to deploy stochastic methods at scale. We demonstrate the capabilities of Korali for Bayesian inference and optimization studies using existing high-performance software such as LAMMPS (CPU-Based) and Mirheo (GPU-Based) and show scaling efficiently on up to 4096 nodes of the CSCS Piz Daint supercomputer.
READ FULL TEXT
