ISBN 978-3-8439-3586-9

978-3-8439-3586-9, Reihe Informatik

Sebastian Rettenberger
Scalable I/O on Modern Supercomputers for Simulations on Unstructured Meshes

142 Seiten, Dissertation Technische Universität München (2018), Softcover, A5

Zusammenfassung / Abstract

Accessing large datasets on persistent storage (I/O) is evolving into one of the major challenges in peta- and exascale computing. The current trend shows an increasing gap between computing power and I/O bandwidth. However, many scientific applications write huge datasets to permanent storage for later visualization and analysis or for checkpointing. Additionally, applications load high resolution datasets at initialization to compute detailed and accurate results.

There are many existing I/O libraries for high-performance computing from low-level libraries (e.g., MPI-IO) to high-level libraries such as HDF5, SIONlib or ADIOS. These libraries are designed to work on parallel file systems, and increase the performance of parallel I/O or help to store data in a structured format. However, the flexibility that arises from the different libraries, file systems, and tuning parameters puts a huge burden on application developers. Selecting the best combination of I/O libraries and tuning parameters is a challenging task. Especially the inconvenient I/O patterns of unstructured datasets can lead to a suboptimal performance.

This thesis bridges a gap between existing I/O libraries and scientific applications with a focus on the massively parallel application SeisSol. Specialized libraries for large unstructured datasets are developed to maximize the I/O bandwidth and to allow efficient usage of modern supercomputers. Optimizations target the complete simulation pipeline and range from reading large meshes to writing results and checkpointing. An additional asynchronous I/O implementation can hide I/O operations similar to asynchronous node-to-node communication by utilizing special I/O cores. Exploiting the optimized I/O stack, the first realistic seismic simulations at petascale level are conducted in this thesis, including a 3D visualization of the results. These simulations are based on unstructured meshes with more than 200 million elements and 100 billion degrees of freedom.