Title: Optimization on ARM architecture of the massively parallel Gysela code dedicated to the simulation of Tokamak fusion plasma
Date: 2019-12-16 15:51
Slug: job_d486d26010ebdb6417d54aa7fe91bd1c
Category: job
Authors: Mathieu Lobet
Email: mathieu.lobet@cea.fr
Job_Type: Post-doctorat
Tags: postdoc
Template: job_offer
Job_Location: Maison de la Simulation
Job_Duration: 1 an
Job_Website: http://www.maisondelasimulation.fr/emploi/1-year-postdoc-position-optimization-on-arm-architecture-of-the-massively-parallel-gysela-code-dedicated-to-the-simulation-of-tokamak-fusion-plasma/
Job_Employer: CEA
Expiration_Date: 2020-04-01
Attachment: job_d486d26010ebdb6417d54aa7fe91bd1c_attachment.pdf
Context
From early 2019, Institute for Magnetic Fusion Research at CEA (Commissariat à l’Energie Atomique et aux Energies Alternatives) is associated to the European project EoCoE-II to optimize and prepare the scientific application Gysela to forthcoming exascale infrastructure. The Gysela code models the electrostatic branch of the Ion Temperature Gradient turbulence in tokamak plasmas. It has been developed at CEA in collaboration with several partners (INRIA, University of Strasbourg, University of Nancy, IPP Garching).
The general purpose of EoCoE, the Energy Oriented Center of Excellence, is to provide support to HPC (High Performance Computing) applications that belong to scientific communities having potential impact on technologies linked to energy. Coordinated by CEA, the consortium gathers 20 partners from 6 European countries and has developed a structured support strategy to enhance the computing performance of selected energy-oriented applications. EoCoE-II focuses on 5 energy Scientific Challenges (Wind, Meteo, Materials, Water, Fusion) that will be supported by 4 Technical Challenges (solvers, programming models, IO and data flow, ensemble runs). Each Scientific Challenge is constituted of a flagship simulation code with a limited number of satellite codes. The main project objectives intends to redesign, develop and optimize flagship community applications for forthcoming generation of super-computers, and to prepare them for exascale or pre-exascale machines. Gysela is the flagship code of the Fusion prospective area selected in this project, with the aim of being exascale-ready in 2 years. The new code will be renamed GyselaX.
ARM processors is one of the possible technology that may equip future exascale super-computers in Europe and in Japan. The Riken Center for Computational Science (RCCS) in Kobe (Japan) is currently developing a first exascale super-computer called Fugaku using Fujitsu ARM processors (ARM8.2-A) composed of 48 vector cores. In Europe, the Mont-Blanc project aims at demonstrating the capability of ARM-based processors for achieving high-performance computing and potentially equip future European exascale super-computers. The RCCS has recently partnered with CEA to work on porting applications toward this kind of technologies.
The EoCoE Center of Excellence will finance this one-year position. The postdoctoral researcher will benefit from a close collaboration with RCCS in Japan to adapt Gysela on the Fugaku super-computer during the first period. It should be noted that extending this collaboration beyond its original duration is a realistic perspective, with the possibility to work at RCCS.
Objectives
The main objective of this position is to adapt and optimize GyselaX for future ARM-based super-computers. It can be divided as follows:
- Ensure that GyselaX can compile and run on ARM processors dedicated to HPC
- Optimize GyselaX to run efficiently on ARM nodes.
- Improve vectorization and memory management of the code for future ARM vector processors (Cavium ThunderX4 and ARM A64fx)
- Address potential scalability issue on extremely large number of nodes by redesigning the exchange mechanism and exploit asynchronism
If the collaboration with Japan is possible:
- Adapt the code to make it run on a node of the post-K Fugaku super-computer in Japan and take advantage of the on-package high-bandwidth memory
- Contribute to improve the Fugaku super-computer environment by interacting closely with Riken HPC experts
- Contribute to demonstrate the computational power of the Fugaku super-computer on fusion applications
Required skills
- PhD or master's degree in a scientific domain strongly connected to HPC
- Operational knowledge of techniques and programming language (Fortran90, C, C++) for application development
- Experience in application parallelization (MPI, OpenMP) and scientific codes optimization on various architectures (SMP, MPP) running in Unix environment
- Skills to work in English
- Knowledge of scripting languages (Python, bash, …) and Unix environment
Location
The successful candidate will work at Maison de la Simulation near Saclay in France. She/He will also work closely with the development team of Gysela located at CEA Cadarache. Regular meetings will be organized to synchronize with them. The candidate will also benefit from the expertise of CINES at Montpellier, a research unit that is currently contributing to the optimization work of Gysela. In the event of a deeper collaboration with RCCS, the candidate may have the opportunity to travel to Japan and work with the local teams.