accelerating wrf i/o performance with adios2 and network-based streaming

Erick Fredj, Yann Delorme, Sameeh Jubran, Mark Wasserman, Zhaohui Ding,Michael Laufer

CoRR(2023)

引用 0|浏览13
暂无评分
摘要
With the approach of Exascale computing power for large-scale High Performance Computing (HPC) clusters, the gap between compute capabilities and storage systems is growing larger. This is particularly problematic for the Weather Research and Forecasting Model (WRF), a widely-used HPC application for high-resolution forecasting and research that produces sizable datasets, especially when analyzing transient weather phenomena. Despite this issue, the I/O modules within WRF have not been updated in the past ten years, resulting in subpar parallel I/O performance. This research paper demonstrates the positive impact of integrating ADIOS2, a next-generation parallel I/O framework, as a new I/O backend option in WRF. It goes into detail about the challenges encountered during the integration process and how they were addressed. The resulting I/O times show an over tenfold improvement when using ADIOS2 compared to traditional MPI-I/O based solutions. Furthermore, the study highlights the new features available to WRF users worldwide, such as the Sustainable Staging Transport (SST) enabling Unified Communication X (UCX) DataTransport, the node-local burst buffer write capabilities and in-line lossless compression capabilities of ADIOS2. Additionally, the research shows how ADIOS2's in-situ analysis capabilities can be smoothly integrated with a simple WRF forecasting pipeline, resulting in a significant improvement in overall time to solution. This study serves as a reminder to legacy HPC applications that incorporating modern libraries and tools can lead to considerable performance enhancements with minimal changes to the core application.
更多
查看译文
关键词
streaming,adios2,performance,network-based
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要