Currently job artifacts in CI/CD pipelines on LRZ GitLab never expire. Starting from Wed 26.1.2022 the default expiration time will be 30 days (GitLab default). Currently existing artifacts in already completed jobs will not be affected by the change. The latest artifacts for all jobs in the latest successful pipelines will be kept. More information:

Commit 0f1846da authored by Dominic Etienne Charrier's avatar Dominic Etienne Charrier
Browse files

Add changelog which I will use tomorrow to update the guidebook.

parent 1cbd5404
* Move global master/worker communication out of Peano's adapters during time stepping iterations.
(Keep it in Peano adapters during mesh refinement iterations.)
* There are now two hook (beginTimeStep,endTimeStep) for user solvers which can be used to
initialise or shut down file operations and to perform global reductions and
broadcasts (among working ranks).
* Const-correctness established for solver macro kernels and USER PDE kernels
Changes applied to all specification files found in ExaHyPE-Engine,
ExaHyPE-Astrophysics and ExaSeis.
* Profiling: I defined Intel ITAC user events for all solvers.
This should give a good idea where we encounter
scalability bottlenecks.
Global Optimisation Parameters
New options:
* limiting = <enum>, <enum> in ["static","dynamic"]
* mesh_refinement = <enum>, <enum> in ["static","dynamic"]
Both are used by the solvers to skip certain operations such as FV-DG
projections in troubled FV cells, the evaluation of DMP and PAD
during the time stepping iterations, the evaluation of the refinement
criterion during time stepping iterations, ...
Modified options:
* fuse_algorithmic_steps = <bool> -> fuse_algorithmic_steps = <enum>, <enum> in ["all","most","none"]
"most" fuses two out of the three ADER-DG phases, "all" fuses all three.
"all" might require reruns when a time step size estimate was wrong.
No reruns are required when a linear scheme is used.
* fuse_algorithmic_steps_factor = <number> -> fuse_algorithmic_steps_rerun_factor = <number>
-> fuse_algorithmic_steps_diffusion_factor = <number>
If a rerun is necessary, the time step size is reset to
dt = dt_cfl * fuse_algorithmic_steps_rerun_factor
Otherwise, the next time step size is estimated as
dt = 0.5 * ( dt_cfl * fuse_algorithmic_steps_diffusion_factor + dt_old)
(diffusion factor as a smaller time step size adds additional numerical diffusion.)
* Nonfused variants "all and "most" do not read and write to the update array anymore in compute cells.
This should increase arithmetic intensity.
* Update array of compute cells is only read and written when necessary.
It is not accessed anymore during batched "all" fused timed stepping iterations.
This should increase arithmetic intensity.
* Completely cleaned the time stepping procedures. Got rid of the predictor
time stamp and step size on each patch.
The refactored code was tested with TestCases/AMR_bug and
TestCases/nonlinear_Euler_Sod_3d, using optimised and generic kernels.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment