Cluster Scalability of Implicit and Implicit-Explicit LS-DYNA Simulations Using a Parallel File System
The parallel efficiency and simulation turn-around times of CAE software continue to be an important factor behind engineering and scientific decisions to develop models at higher fidelity. Most parallel LS-DYNA simulations use scalable Linux clusters for their demanding HPC requirements, but for certain classes of FEA models, data IO can severely degrade overall scalability and limit CAE effectiveness. As LS-DYNA model sizes grow and the number of processing cores are increased for a single simulation, it becomes critical for each thread on each core to perform IO operations in parallel, rather than rely on the master compute thread to collect each IO process in serial. This paper examines the scalability characteristics of LS-DYNA for implicit and implicit-explicit models on up to 256 processing cores. This joint study conducted by the University of Cambridge and Panasas, used an HPC cluster environment that combines a 28 TFLOP Intel Xeon cluster with a Panasas shared parallel file system and storage. Motivation for the study was to quantify the performance benefits of parallel I/O in LS-DYNA for large-scale FEA simulations on a parallel file system vs. performance of a serial NFS file system. The LS-DYNA models used for the study comprise cases that were relevant in size and physics features to current LSTC customer practice. The favourable results demonstrate that LS-DYNA with parallel I/O will show significant benefit for advanced implicit simulations that can be heavy in I/O relative to numerical operations. These performance benefits were shown to extend to a mix of concurrent LS-DYNA jobs that require concurrent data writes to a shared file system, which for an NFS-based file system would still bottleneck from its single data path for I/O. The paper also reviews CAE workflow benefits since, as an LS-DYNA simulation is completed, the same shared storage provides a platform for direct post-processing and visualization without the need for large file transfers.
https://www.dynalook.com/conferences/copy_of_european-conf-2009/N-II-04.pdf/view
https://www.dynalook.com/@@site-logo/DYNAlook-Logo480x80.png
Cluster Scalability of Implicit and Implicit-Explicit LS-DYNA Simulations Using a Parallel File System
The parallel efficiency and simulation turn-around times of CAE software continue to be an important factor behind engineering and scientific decisions to develop models at higher fidelity. Most parallel LS-DYNA simulations use scalable Linux clusters for their demanding HPC requirements, but for certain classes of FEA models, data IO can severely degrade overall scalability and limit CAE effectiveness. As LS-DYNA model sizes grow and the number of processing cores are increased for a single simulation, it becomes critical for each thread on each core to perform IO operations in parallel, rather than rely on the master compute thread to collect each IO process in serial. This paper examines the scalability characteristics of LS-DYNA for implicit and implicit-explicit models on up to 256 processing cores. This joint study conducted by the University of Cambridge and Panasas, used an HPC cluster environment that combines a 28 TFLOP Intel Xeon cluster with a Panasas shared parallel file system and storage. Motivation for the study was to quantify the performance benefits of parallel I/O in LS-DYNA for large-scale FEA simulations on a parallel file system vs. performance of a serial NFS file system. The LS-DYNA models used for the study comprise cases that were relevant in size and physics features to current LSTC customer practice. The favourable results demonstrate that LS-DYNA with parallel I/O will show significant benefit for advanced implicit simulations that can be heavy in I/O relative to numerical operations. These performance benefits were shown to extend to a mix of concurrent LS-DYNA jobs that require concurrent data writes to a shared file system, which for an NFS-based file system would still bottleneck from its single data path for I/O. The paper also reviews CAE workflow benefits since, as an LS-DYNA simulation is completed, the same shared storage provides a platform for direct post-processing and visualization without the need for large file transfers.
N-II-04.pdf
— 350.7 KB
