-
Notifications
You must be signed in to change notification settings - Fork 189
Proceedings 2024 ESPResSo meetings
Jean-Noël Grad edited this page May 14, 2024
·
13 revisions
- LB GPU now works in parallel
- CUDA-aware MPI is still a work-in-progress
- work on multi-GPU support has just started
- long-term plan: multi-GPU support with 1 GPU per MPI rank and multiple shared memory threads per MPI rank via OpenMP
- multi-system simulations are now possible for almost all features
- caveats: two systems cannot have particles with the same particle ids, Monte Carlo not yet supported
- can be enabled with a one-liner change to
system.py(see last commit in jngrad/multiverse)
- convert particle cells from AoS to SoA (#4754), i.e. one array per particle property
- improves cache locality and CPU optimizations
- use Cabana to hide optimizations
- bump all version requirements (#4905)
- on ICP workstations, only need to update formatting and linter tools with
pip3 install -r requirements.txt autopep8 pycodestyle pylint pre-commit
- ModEMUS project: nanoparticle diffusion in hydrogel network, by Pablo Blanco (NTNU) (see Ma et al. 2018)
- currently implemented with Langevin, plan is to use LB instead to improve accuracy
- ultrasound streaming could be modeled with a gradient pressure via pressure boundary conditions
- GPU LB with particle coupling implemented in #4734
- requires CUDA>=12.0 to make double-precision
atomicAdd()available - performance is degraded when using more than 1 MPI rank to communicate to the same GPU, need to look into CUDA-aware MPI
-
pyMBE-dev/pyMBE now uses EESSI for CI/CD
- continuous integration PR: pyMBE-dev/pyMBE#1
- continuous delivery PR: pyMBE-dev/pyMBE#23
- deployed docs: https://pymbe-dev.github.io/pyMBE
- Thursday: HPC team meeting to discuss software stack
- Monday: set up the software stack
- Tuesday: online meeting with the company
- main objectives of the CoE MultiXscale:
- EESSI: "app store" for scientific software
- multiscale simulations with 3 pilot cases: helicopter blades turbulent flow, ultrasound imaging of living tissues, energy storage devices
- make software pre-exascale ready
- training on using these software
- ongoing projects for the ICP:
- ZnDraw+SiMGen project in collaboration with Gábor Csányi
- demo available at https://zndraw.icp.uni-stuttgart.de/
- LB boundaries in the waLBerla version of ESPResSo are broken when using 2 or more MPI ranks
- the LB ghost layer doesn't contain any information about boundaries, so fluid can flow out into the neighboring LB domain, where it gets trapped into the boundary
- more details can be found in the bug report #4859
- the solution will require a ghost update after every call to a node/slice/shape boundary velocity setter function
- it is now possible to choose the exact equations of motions solved for each particle
- more combinations of integrators and thermostats are allowed
- one can mix different types of virtual sites in the same simulation
- a new Python interface needs to be designed (target for next meeting)
- CUDA 12.0 is now the default everywhere at the ICP
- GTX 980 GPUs are being removed from ICP workstations (GTX 1000 and higher are needed for double-precision
atomicAdd) - Python 3.12 changes the way we use the unittest module (#4852)
- Python virtual environments becomes mandatory for
pip installin Ubuntu 24.04- user guide needs to reflect that change
- ZnDraw bridge is currently being developed for ESPResSo
- supports live visualization in the web browser
- the plan is to re-implement all features available in the ESPResSo OpenGL visualizer
- CUDA 12 will be the default in Ubuntu 24.04, due April 2024
- CUDA 12 is required for C++20, which makes contributing to ESPResSo significantly easier (#4846)
- compute clusters and supercomputers where ESPResSo is currently used already provide compiler toolchains compatible with CUDA 12