FEAST - Finite Element Analysis and Solution Tools
FEAST is our next-generation Finite Element software that implements many ideas and concepts of hardware-oriented numerics. FEAST is explicitly designed for large-scale runs on supercomputers. Its core features include:
- All linear systems are solved by a hybrid non-overlapping domain decomposition parallel multilevel/multigrid method called ScaRC, short for Scalable Recursive Clustering. ScaRC works on what we call globally unstructured locally structured grids to achieve both geometric flexibility and scalability at large and maximum local performance. All local problems are structured, and this structure is exploited in the design of fast local multigrid solvers that achieve a substantial fraction of the available peak performance.
- FEAST is able to exploit GPUs as co-processors, which happens to be one of the results of my PhD research.
- Applications exist for solid mechanics and fluid dynamics.
More information, including a detailed bibliography, is available on the project page. As FEAST still undergoes massive changes (like most actively developed research codes), we haven't made a public release yet, simply because we do not have the time to support it. This may change in the future.
SPECFEM is a family of codes for wave propagation modeling, e.g. in seismics and acoustics. Various specialised versions exist, e.g. in 1D, 2D and 3D, the latter for production runs either for sedimentary basins (full flexibility) or at the scale of continents and the Earth (full performance). Lots of geological models and effects like coriolis forces, PML conditions, attenuation etc. are included. Numerically, the software is based on explicit integration in time combined with a spectral element discretisation in space, which ensures almost pseudo-spectral accuracy while retaining all the benefits from a classical finite element approach. SPECFEM3D_GLOBE has been shown to scale extremely well in parallel (both in the weak and strong sense) and to achieve substantial fractions of peak performance. I have been honoured to contribute to the GPU-accelerated version of the code, and to the port for ARM architectures.
TOMOFAST3D is a code that Roland Martin, Dimitri Komatitsch and I are currently working on. As the name suggests, it aims at quasi-real-time imaging / tomography, using small clusters of accelerator-equipped PCs as a target platform, in particular NVIDIA GPUs and Intel Xeon Phi's. We are currently investigating capacitance tomography for two-phase flows (e.g. oil with gas bubbles), using specialised multigrid schemes for the forward problems and a combination of wavelets and LSQR for the inverse imaging step.
The code will be made publicly available once it has reached a stable state and we have published an accompanying paper.
HONEI - Hardware-Oriented Numerics Efficiently Implemented
HONEI aims at developing a set of heavily templated C++ libraries to facilitate the development of scientific software on heterogeneous architectures. The underlying idea is to abstract from the hardware as much as possible, and ease the implementation of user-provided application-specific kernels where unavoidable. HONEI provides a rich set of infrastructure for multicore CPUs, GPUs, the Cell processor and MPI-based clusters. Please check the project web page for details.
DeViSoR GRID3D is an effort to implement a FEM preprocessor, including mesh generation, geometry definition etc. in an intuitive point-and-click vector drawing GUI. This software is written entirely in Java. I used to work a lot on this project, in particular for my diploma thesis. You can download a fully documented version including source code and manuals from the corresponding project page.
FeatFlow is a scientific code for 2D and 3D stationary and instationary Navier-Stokes simulations. I contributed with my diploma thesis (some new mesh smoothers and a geometry projection scheme) and wrote parts of the GNU-make based installer for the current version 1.3. FeatFlow is implemented in good-old sequential Fortran 77. FeatFlow is available for download free of charge, including the source code, a library of results, and tons of other information.