In high-performance computing (HPC), efficient data distribution across multi-node systems is crucial for optimizing parallel processing. Traditional static data distribution strategies often fall short in dynamic workloads, leading to suboptimal performance. This thesis investigates the Reshuffle library, which introduces elastic data distribution capabilities at runtime, leveraging the Message Passing Interface (MPI) to dynamically adjust data layouts across nodes. The study focuses on integrating the Reshuffle library within a distributed heat transfer simulation framework, using the Jacobi method as the core algorithm due to its simplicity and parallel efficiency. The research employs a modular system architecture to facilitate the integration of the Reshuffle library, enabling dynamic data redistribution and enhancing load balancing and resource utilization. A series of benchmarking experiments were conducted to evaluate the library’s performance across varying grid sizes and MPI rank configurations. The study also highlights the impact of different layout configurations on reshuffling efficiency. This thesis contributes to the field of elastic HPC frameworks by providing a practical implementation and assessment of the Reshuffle library, offering insights into its potential applications beyond heat transfer simulations. Future work could explore the library’s integration with other computational frameworks and optimize its algorithms for broader applicability.     «

In high-performance computing (HPC), efficient data distribution across multi-node systems is crucial for optimizing parallel processing. Traditional static data distribution strategies often fall short in dynamic workloads, leading to suboptimal performance. This thesis investigates the Reshuffle library, which introduces elastic data distribution capabilities at runtime, leveraging the Message Passing Interface (MPI) to dynamically adjust data layouts across nodes. The study focuses on integra...     »