Scalable Multi-Platform PIC Simulations as an Open Science Service

PIConGPU is a fully open, community-driven, 3D and 2D3V particle-in-cell code for the age of heterogeneous, many-core driven supercomputing. Developed in a single source C++ code base, PIConGPU supports both "traditional" CPU architectures as well as modern and highly parallel architectures such as OpenPOWER, Xeon Phi, and Nvidia GPUs.

PIConGPU has shown to be suitable for production runs on the full system size of TOP5 clusters such as Titan (ORNL) and Piz Daint (CSCS). Machines like those enable few-hour turnarounds for full 3D3V simulations on complex studies such as laser-ion acceleration from mass-limited targets, long-scale laser-wakefield acceleration with high bunch charges, and hybrid acceleration schemes. The resulting output of systematic parameter scans (PBytes+) raises a severe challenge for data centers. We address these issues with modern IO frameworks, performance modeling, and in situ data reduction techniques. Using such online methods we can investigate a wide range of observables relevant for experiments and run dozens of simulations at the same time frame as an experimental beam time.

PIConGPU is further complemented by modern methods for photon generation, transport, as well as X-ray interaction. This simulation framework aims to provide documented, installable, and re-usable software components for the community, well-suited for open data (openPMD) and open science workflows without restrictions. Latest developments include a python-centric, extensive framework for specific experiments, which provides all of the above in an intuitive, non-expert user interface.