Experiences with Porting the FLASH Code to Ookami, an HPE Apollo 80 A64FX Platform
Experiences with Porting the FLASH Code to Ookami, an HPE Apollo 80 A64FX Platform
复制标题
DOI：
10.1145/3503470.3503478
复制
发表时间：
2022-01
期刊：
影响因子：
--
通讯作者：
Catherine Feldman;Benjamin Michalowicz;E. Siegmann;Tony Curtis;A. Calder;R. J. Harrison
中科院分区：
文献类型：
--
作者：
Catherine Feldman;Benjamin Michalowicz;E. Siegmann;Tony Curtis;A. Calder;R. J. Harrison
我们介绍了在Ookami上运行社区模拟代码FLASH的初步经验，该代码由芝加哥大学开发，用于多尺度多物理应用，Ookami是一个采用富士通开发的A64FX处理器的技术测试平台。我们的工作主要集中在运行FLASH“开箱即用”，看看哪些编译器和软件实现（例如MPI）的组合允许代码以最小的修改运行。FLASH是部署Ookami的一个大项目；它是对新安装软件的不同版本的测试，也是Ookami网站FAQ页面的基础。我们用不同的编译器和其他软件报告我们的结果，以及我们最初的缩放结果，并尝试利用A64FX的SVE指令和NUMA架构。我们发现FLASH很容易与不同的编译器和MPI实现一起运行，并且显示出预期的良好可伸缩性。然而，要充分利用A64FX的架构特点，并为Ookami上的FLASH产生显著的加速，还需要做更多的工作。
We present initial experiences with running the community simulation code FLASH, developed at the University of Chicago for multi-scale multi-physics applications, on Ookami, a technology testbed featuring the A64FX processor developed by Fujitsu. Our effort focused largely on running FLASH “right out of the box” to see which combinations of compilers and software implementations (e.g. MPI) allowed the code to run with minimal modification. FLASH was one application in a larger effort to deploy Ookami; it served as a test for different versions of newly installed software, and as a cornerstone for the FAQ page of the Ookami website. We report on our results with different compilers and other software, along with our initial scaling results and attempts to utilize the A64FX’s SVE instructions and NUMA architecture. We found that FLASH readily ran with different compilers and MPI implementations, and showed the expected good scaling with no turning. However, more work must be done to fully take advantage of the A64FX’s architectural features and produce a significant speedup for FLASH on Ookami.