Identifying Degree and Sources of Non-Determinism in MPI Applications Via Graph Kernels

As the scientific community prepares to deploy an increasingly complex and diverse set of applications on exascale platforms, need assess reproducibility simulations identify root causes failures increases correspondingly. One greatest challenges facing issues at is inherent non-determinism level inter-process communication. The use non-deterministic communication constructs necessary boost performance, but can also hamper software correctness result reproducibility. To address this challenge, we propose a framework for identifying percentage sources non-determinism. We model parallel executions as directed graphs leverage graph kernels characterize run-to-run variations in demonstrate effectiveness kernel similarity proxy non-determinism, by showing that these quantify type degree present patterns. our framework's ability link runtime sources, with adaptive mesh refinement application, where automatically quantifies impact function calls Monte Carlo parameter configurations