Extreme Scale Computing

Advanced Algorithms and Frameworks for Extreme Scale Computing

This WP focuses on advanced algorithms and frameworks for extreme scale computing. The team will implement and study the applicability of the innovative task-based framework Eventify in the simulation code SMILEI (www.maisondelasimulation.fr/categorie-actualites/smilei) for asynchronous execution on heterogeneous systems. The team is facing several numerical and software design challenges:

• benefiting from efficient asynchronism on CPUs,
• extending the task-based method to GPUs,
• being able to use all computing resources on heterogeneous nodes.

On FZJ side, the team will focus on developing an event-driven uniform task-based programming framework for both CPUs and GPUs (Eventify), and on CEA side, the team will extend and strengthen the parallel algorithm of the Particle In Cell (PIC) code SMILEI for running on CPU and GPU exascale machines.

Besides theory and experiment, today’s scientific research is heavily driven by simulations. However, such simulations demand a lot of computing resources and often even require top-class supercomputers to enable scientific progress. Therefore, we are focusing on developing, porting and testing of advanced parallel software for such systems. Since parallel programming is not trivial and keeping all processors busy all the time needs careful implementations, we investigate new programming models for current and future heterogeneous HPC hardware architectures. This allows us to exploit the already available hierarchical hardware parallelism even better, while still keeping the simulation software flexible for methodical extensions. The appropriate programming model can also help to reduce complexity for the programmer and mitigate conflicting requirements like efficiency, portability and maintainability.

Specifically, within AIDAS, we are studying modern task-based programming models beyond classical fork-join schemes. The available, massive hardware parallelism of our supercomputers demands fine-grained tasks with low-overhead to utilize such clusters to the fullest. Modern C++ and lightweight synchronization primitives can help us to pave the road for future scientific software to support cluster nodes with hundreds of CPUs and GPU cores efficiently. As a demonstrator, we explore the use of task-based parallelism on the Particle-In-Cell method used to simulate plasma physics scenarii (laser-matter interaction, astrophysics and more). Our demonstrator is a mini-app based on the parallel simulation application Smilei partly developed at Maison de la Simulation (CEA Saclay).