Accelerating Magnetospheric Modeling with Heterogeneous Hardware
Abstract
The advent of supercomputers with multiple accelerators per computational node is impacting the development of codes for magnetospheric modeling. The two current largest supercomputers in November 2018 Top500 list, Summit and Sierra, feature six and four V100 NVIDIA GPUs per node respectively providing a theoretical peak performance of 750 and 500 Tops/s (operations in mixed precision) per node [1]. However, it is still unclear how codes for magnetospheric modeling could take advantage of new heterogeneous architectures. Widely-used massively parallel codes for magnetospheric modeling are not yet capable of exploiting these new systems and need to be redesigned. Two main aspects have to be considered: first, the algorithms for magnetospheric modeling have to be reformulated to use dense matrix operations; second, new algorithms have to cope with low-precision calculations, still retaining acceptable accuracy. In this talk, we review emerging heterogeneous architectures and present our work in designing and developing new algorithmic changes in iPIC3D [2], an implicit PIC code for magnetospheric modeling, to fully exploit heterogeneous hardware.
This work has received funding from the European Commission H2020 program, Grant Agreement No. 801039 (EPiGRAM-HS, epigram-hs.eu) [1] Markidis, Stefano, Steven Wei Der Chien, Erwin Laure, Ivy Bo Peng, and Jeffrey S. Vetter. "NVIDIA tensor core programmability, performance & precision." In 2018 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pp. 522-531. IEEE, 2018. [2] Markidis, Stefano, Giovanni Lapenta, and Rizwan-uddin. "Multi-scale simulations of plasma with iPIC3D." Mathematics and Computers in Simulation 80, no. 7 (2010): 1509-1519.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMSM12B..07M
- Keywords:
-
- 1910 Data assimilation;
- integration and fusion;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS;
- 2753 Numerical modeling;
- MAGNETOSPHERIC PHYSICS;
- 7924 Forecasting;
- SPACE WEATHER