-
Notifications
You must be signed in to change notification settings - Fork 199
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Start: PerformanceHints After Init #1630
Start: PerformanceHints After Init #1630
Conversation
Start a helper routine that gives performance hints after initialization of the simulation.
1c900a8
to
9f204ef
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks, looks good!
{ | ||
// Check requested MPI ranks and available boxes | ||
amrex::Long total_nboxes = 0; // on all MPI ranks | ||
for (int ilev = 0; ilev <= finestLevel(); ++ilev) { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The question was raised in Tuesday morning's meeting about whether work is done by level or all levels at once. If it is by level, this should use the number of boxes only on level 0. Correct?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This is checking all levels right now, assuming that we can distribute refined levels to otherwise idle ranks that don't have a box from level 0 to crunch on.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
That was my question, is that a good assumption? Did someone check it?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, I think that this is a good assumption, at least for the starting configuration (i.e. without explicit load-balancing).
For instance, using the above example, let us assume that we have 4 MPI rank (numbered from 0 to 3) and 3 boxes on level 0 + 1 box on level 1.
In that case, amrex
will create the 3 boxes from level 0 on ranks 0,1,2. It will then create the box for level 1, and automatically place it on level 1 on rank 3, because amrex
uses hints on the memory usage of each MPI ranks, when allocating new boxes.
At least that is my understanding, and I did check this a few months ago. But maybe @WeiqunZhang can confirm.
total_nboxes += boxArray(ilev).size(); | ||
} | ||
if (ParallelDescriptor::NProcs() > total_nboxes) | ||
amrex::Print() << "\n[Warning] [Performance] Too many resources / too little work!\n" |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Could this be printed in color? You can use the escape sequences that set the color.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hehe, good question.
In practice, escape sequences are sometimes terminal/platform dependent and we might want to add options for this (enable, disable), because not all users know to open files with escape sequences with less -R
for instance.
Some terminal settings also highlight keywords like error or warning by default.
Then there is the question of light vs. dark background and good color schemes.
One could use a library like {fmt} for reasonable color support and excellent pythonic formatting. Worth considering because colors are great fun, but I would not do it in this PR yet.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Trying to think of a way to make the message more noticeable since it will quickly get lost in the output.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@dpgrote I agree: this is a recurring discussion that we have (e.g. this was also discussed for unused parameters).
Alternatives that were mentioned before include aborting immediately (a bit exaggerated in this particular case), printing the message at each timestep (maybe an issue if we have many such warnings...) and including the message in a separate file (but the user may not necessarily see this).
Any other suggestion is definitely welcome, since the above ones are not ideal.
That being said, it is probably a larger issue than that particular PR, so I think it is fine to merge this one in its current state.
…rpX#1624) * Fix bug in momentum-conserving interpolation function * Reset checksum benchmark for CI test momentum-conserving-gather Unused Params: Check after Step 0 (ECP-WarpX#1596) An early check for unused parameters after the first step. Add Superparticle version of getParticlePosition. (ECP-WarpX#1640) * Add Superparticle version of GetParticlePosition. * move unpack_particle to a free function and rename * no longer need SuperPType * Update Source/Particles/Pusher/GetAndSetPosition.H * remove templating from get_particle_position * remove template * Add missing include Co-authored-by: Axel Huebl <axel.huebl@plasma.ninja> Distribution mapping and cost plotting (ECP-WarpX#1444) * Distribution mapping and cost plotting Cleanup Cleanup Cleanup EOL Unused import * Cost initializes to 0.0 * plot slices of 3D * WIP * WIP docs * docs * docs * docs * docs * docs * docs * docs * docs * EOL SpeciesProperties: C, N, O, Cu (ECP-WarpX#1638) * SpeciesProperties: C, N, O, Cu Add more ionic species as pre-defined particle species. * Cu: Add to Ionization Energies Fix 2D: ParticleHistogram (ECP-WarpX#1635) Fix out-of-bounds access in particle histogram in non-3D. Fix FilterFunctor in non-3D (ECP-WarpX#1633) Fix an out-of-bounds access to positions in 2D & RZ for filter functors. openPMD: ionizationLevel (ECP-WarpX#1622) * openPMD: ionizationLevel Write out the ionizationLevel with openPMD. * openPMD Record Repetitions: Use Lambdas openPMD: Particle Filter (Container) (ECP-WarpX#1632) * openPMD: Particle Filter (Container) A fresh implementation of particle filters for openPMD diagnostics using the generalized `PhysicalParticleContainer` approach that we also use in plotfiles 🎉 * Ion Example: Filter & Coarsen Test diagnostics filter & coarsen functions in CI. Test reduced diagnostics (histograms). Define: _OPENMP -> AMREX_USE_OMP (ECP-WarpX#1520) * Define: _OPENMP -> AMREX_USE_OMP Replace the define check of `_OPENMP` with the explicit backend control of `AMREX_USE_OMP` for parallel constructs. Doing so avoids that we accidentially turn on OpenMP, e.g. if a dependency pulls it in for linear algebra, I/O, etc. This can led to confusion if the user explicitly requested a serial build. Also, we might want to use OpenMP functionality here and there for auxiliary functions w/o having to use the AMReX OpenMP backend, i.e. because we compile for GPUs. * Add missing amrex::Gpu::notInLaunchRegion Start: PerformanceHints After Init (ECP-WarpX#1630) Start a helper routine that gives performance hints after initialization of the simulation. Fix number of guard cells in PML with 2D PSATD (ECP-WarpX#1648) Remove unused lines about particle filters (ECP-WarpX#1646) Add parameter for default galilean velocity (ECP-WarpX#1097) * Add parameter for default galilean velocity * Apply suggestions from code review Co-authored-by: Edoardo Zoni <59625522+EZoni@users.noreply.github.com> * Fix bug in constructor of PhysicalParticleContainer * Use new input parameter in CI test Co-authored-by: Edoardo Zoni <59625522+EZoni@users.noreply.github.com> Co-authored-by: Edoardo Zoni <ezoni@lbl.gov> Cleaned up field diagnostic in picmi interface (ECP-WarpX#1649) Use pinned memory for tmp particles in diags. (ECP-WarpX#1644) * Use pinned memory for tmp particles in diags. * openPMD: pinned memory (tmp particles) Co-authored-by: Axel Huebl <axel.huebl@plasma.ninja> Replace an always false test with an abort (ECP-WarpX#1655) * replace always false test with abort * fix bug [mini-PR] Broken link to scripts (ECP-WarpX#1651) CI: Ubuntu-latest -> 20.04 (ECP-WarpX#1654) Start Embedded Boundary development (ECP-WarpX#1641) * Start Embedded Boundary development Modify the build system for embedded boundary (EB) support. Currently EB is a compile time option that must be explicitly enabled. For GNU Make, one can enable it with `USE_EB=TRUE`, whereas for CMake, `-DAMReX_EB=ON`. Later we could decide to enable EB by default with all regular geometry. Add a simple geometry initialization function, WarpX::InitEB. By default, the geometry is all regular. A few basic types such as box, cylinder, plane, sphere, etc. are supported via ParmParse runtime parameters. See `amrex/Src/EB/AMReX_EB2.cpp` for more details. Later, we could build more complex geometry using constructive solid geometry (CSG). ( https://en.wikipedia.org/wiki/Constructive_solid_geometry ) See `amrex/Tutorials/EB/GeometryGeneration` for an example of CSG. There is also a STL ( https://en.wikipedia.org/wiki/STL_(file_format) ) approach under development in AMReX. Add a new member, m_factory, to WarpX class. This object can be used to obtain geometry information such as whether a cell is cut, edge centroids, etc. Currently we are not using these factories to build MultiFabs for field data to embed the geometry information into the data containers. We could do that later if it is needed or it makes things more convenient. Nevertheless, this should be sufficient to start the EB development of the field solver. It's not clear to me yet how many ghost cells are needed for the geometry information. It's currently one, and can be adjusted. In the future, when particle and embedded boundary interaction is considered, we can use `amrex::FillSignedDistance` function to obtain signed distance function on the nodes. With that information, one should be able to determine where and when a particle collides with the embedded boundary. * resize factory vector * CMake: Require AMReX_EB for embedded boundaries * Add AMReX_Config.H et al. Explicit includes are most robust to make sure defines are set with future refactorings. * rename Factory fieldFactory Co-authored-by: Axel Huebl <axel.huebl@plasma.ninja> Add filter for particle histogram (ECP-WarpX#1643) * Add filter for particle histogram * Update benchmarks * Update benchmark again * Use serialize_ics in test + proper usage of RandomEngine * Laser-Ion Acc. (2D3V): Test Hist Filter Co-authored-by: Axel Huebl <axel.huebl@plasma.ninja> Take time step into account to compute guard cells for J and rho (ECP-WarpX#1607) * Use IntVect for ng_J and ng_rho * Compute guard cells for J and rho based on dt * Reset some CI benchmarks * Fix rebase commit * Add back +1 cell for rho: fix remaining out-of-bound accesses * Simplify ASSERTS using new interface of amrex::numParticlesOutOfRange Modern setup.py: CMake-Driven, Multi-Dimensional (ECP-WarpX#1647) This adds a new, project-centric setup.py file. With this file, all dimensions (2D, 3D, RZ) of WarpX can be built and packaged at once, using the CMake build logic. Build & install: ```bash pip wheel -v . pip install *whl ``` Updates for electrostatic solver (ECP-WarpX#1604) * Fix electrostatic solver with momentum conservation * Fix electrostatic, adding call to FillBoundaryAux * For electrostatic, removed unneeded extra calls to UpdateAux and FillBoundary * For electrostatic, calculate fields at the end of the time step * Updated ElectrostaticSphere analysis script to use fields from end of time step Reset broken benchmark for CI test initial_distribution (ECP-WarpX#1661) make sure we redefine the tmp particle tiles when we load balance. (ECP-WarpX#1658) [mini-PR] Add particle tiling documentation (ECP-WarpX#1665) * tiling wip * do_tiling documentation * Update Docs/source/running_cpp/parameters.rst Co-authored-by: Luca Fedeli <luca.fedeli@for.unipi.it> Co-authored-by: Luca Fedeli <luca.fedeli@for.unipi.it> fix inconsistent formatting (ECP-WarpX#1666) parameter.rst fixes (ECP-WarpX#1663) requirements.txt: fix overspecification (ECP-WarpX#1668) I accidentally added the patch-level for version-compatible matching This removes the patch-level for `~=` matching for the packages that have a >=1 major version already. scipy 1.5+ (ECP-WarpX#1670) Sufficient since it was last release in december and works well. The 3.5 release series still builds wheels for Python 3.6 for Ubuntu oldstable (18.04). Remove redundant "do_qed" option in inputfile (ECP-WarpX#1667) * removed redundant do_qed option in inputfile * fixed bug remove semicolons (ECP-WarpX#1662) fixTypo for pml has particles parameter for docs (ECP-WarpX#1671) load balance efficiency reduced diagnostic eol minor eol
…rpX#1624) * Fix bug in momentum-conserving interpolation function * Reset checksum benchmark for CI test momentum-conserving-gather Unused Params: Check after Step 0 (ECP-WarpX#1596) An early check for unused parameters after the first step. Add Superparticle version of getParticlePosition. (ECP-WarpX#1640) * Add Superparticle version of GetParticlePosition. * move unpack_particle to a free function and rename * no longer need SuperPType * Update Source/Particles/Pusher/GetAndSetPosition.H * remove templating from get_particle_position * remove template * Add missing include Co-authored-by: Axel Huebl <axel.huebl@plasma.ninja> Distribution mapping and cost plotting (ECP-WarpX#1444) * Distribution mapping and cost plotting Cleanup Cleanup Cleanup EOL Unused import * Cost initializes to 0.0 * plot slices of 3D * WIP * WIP docs * docs * docs * docs * docs * docs * docs * docs * docs * EOL SpeciesProperties: C, N, O, Cu (ECP-WarpX#1638) * SpeciesProperties: C, N, O, Cu Add more ionic species as pre-defined particle species. * Cu: Add to Ionization Energies Fix 2D: ParticleHistogram (ECP-WarpX#1635) Fix out-of-bounds access in particle histogram in non-3D. Fix FilterFunctor in non-3D (ECP-WarpX#1633) Fix an out-of-bounds access to positions in 2D & RZ for filter functors. openPMD: ionizationLevel (ECP-WarpX#1622) * openPMD: ionizationLevel Write out the ionizationLevel with openPMD. * openPMD Record Repetitions: Use Lambdas openPMD: Particle Filter (Container) (ECP-WarpX#1632) * openPMD: Particle Filter (Container) A fresh implementation of particle filters for openPMD diagnostics using the generalized `PhysicalParticleContainer` approach that we also use in plotfiles 🎉 * Ion Example: Filter & Coarsen Test diagnostics filter & coarsen functions in CI. Test reduced diagnostics (histograms). Define: _OPENMP -> AMREX_USE_OMP (ECP-WarpX#1520) * Define: _OPENMP -> AMREX_USE_OMP Replace the define check of `_OPENMP` with the explicit backend control of `AMREX_USE_OMP` for parallel constructs. Doing so avoids that we accidentially turn on OpenMP, e.g. if a dependency pulls it in for linear algebra, I/O, etc. This can led to confusion if the user explicitly requested a serial build. Also, we might want to use OpenMP functionality here and there for auxiliary functions w/o having to use the AMReX OpenMP backend, i.e. because we compile for GPUs. * Add missing amrex::Gpu::notInLaunchRegion Start: PerformanceHints After Init (ECP-WarpX#1630) Start a helper routine that gives performance hints after initialization of the simulation. Fix number of guard cells in PML with 2D PSATD (ECP-WarpX#1648) Remove unused lines about particle filters (ECP-WarpX#1646) Add parameter for default galilean velocity (ECP-WarpX#1097) * Add parameter for default galilean velocity * Apply suggestions from code review Co-authored-by: Edoardo Zoni <59625522+EZoni@users.noreply.github.com> * Fix bug in constructor of PhysicalParticleContainer * Use new input parameter in CI test Co-authored-by: Edoardo Zoni <59625522+EZoni@users.noreply.github.com> Co-authored-by: Edoardo Zoni <ezoni@lbl.gov> Cleaned up field diagnostic in picmi interface (ECP-WarpX#1649) Use pinned memory for tmp particles in diags. (ECP-WarpX#1644) * Use pinned memory for tmp particles in diags. * openPMD: pinned memory (tmp particles) Co-authored-by: Axel Huebl <axel.huebl@plasma.ninja> Replace an always false test with an abort (ECP-WarpX#1655) * replace always false test with abort * fix bug [mini-PR] Broken link to scripts (ECP-WarpX#1651) CI: Ubuntu-latest -> 20.04 (ECP-WarpX#1654) Start Embedded Boundary development (ECP-WarpX#1641) * Start Embedded Boundary development Modify the build system for embedded boundary (EB) support. Currently EB is a compile time option that must be explicitly enabled. For GNU Make, one can enable it with `USE_EB=TRUE`, whereas for CMake, `-DAMReX_EB=ON`. Later we could decide to enable EB by default with all regular geometry. Add a simple geometry initialization function, WarpX::InitEB. By default, the geometry is all regular. A few basic types such as box, cylinder, plane, sphere, etc. are supported via ParmParse runtime parameters. See `amrex/Src/EB/AMReX_EB2.cpp` for more details. Later, we could build more complex geometry using constructive solid geometry (CSG). ( https://en.wikipedia.org/wiki/Constructive_solid_geometry ) See `amrex/Tutorials/EB/GeometryGeneration` for an example of CSG. There is also a STL ( https://en.wikipedia.org/wiki/STL_(file_format) ) approach under development in AMReX. Add a new member, m_factory, to WarpX class. This object can be used to obtain geometry information such as whether a cell is cut, edge centroids, etc. Currently we are not using these factories to build MultiFabs for field data to embed the geometry information into the data containers. We could do that later if it is needed or it makes things more convenient. Nevertheless, this should be sufficient to start the EB development of the field solver. It's not clear to me yet how many ghost cells are needed for the geometry information. It's currently one, and can be adjusted. In the future, when particle and embedded boundary interaction is considered, we can use `amrex::FillSignedDistance` function to obtain signed distance function on the nodes. With that information, one should be able to determine where and when a particle collides with the embedded boundary. * resize factory vector * CMake: Require AMReX_EB for embedded boundaries * Add AMReX_Config.H et al. Explicit includes are most robust to make sure defines are set with future refactorings. * rename Factory fieldFactory Co-authored-by: Axel Huebl <axel.huebl@plasma.ninja> Add filter for particle histogram (ECP-WarpX#1643) * Add filter for particle histogram * Update benchmarks * Update benchmark again * Use serialize_ics in test + proper usage of RandomEngine * Laser-Ion Acc. (2D3V): Test Hist Filter Co-authored-by: Axel Huebl <axel.huebl@plasma.ninja> Take time step into account to compute guard cells for J and rho (ECP-WarpX#1607) * Use IntVect for ng_J and ng_rho * Compute guard cells for J and rho based on dt * Reset some CI benchmarks * Fix rebase commit * Add back +1 cell for rho: fix remaining out-of-bound accesses * Simplify ASSERTS using new interface of amrex::numParticlesOutOfRange Modern setup.py: CMake-Driven, Multi-Dimensional (ECP-WarpX#1647) This adds a new, project-centric setup.py file. With this file, all dimensions (2D, 3D, RZ) of WarpX can be built and packaged at once, using the CMake build logic. Build & install: ```bash pip wheel -v . pip install *whl ``` Updates for electrostatic solver (ECP-WarpX#1604) * Fix electrostatic solver with momentum conservation * Fix electrostatic, adding call to FillBoundaryAux * For electrostatic, removed unneeded extra calls to UpdateAux and FillBoundary * For electrostatic, calculate fields at the end of the time step * Updated ElectrostaticSphere analysis script to use fields from end of time step Reset broken benchmark for CI test initial_distribution (ECP-WarpX#1661) make sure we redefine the tmp particle tiles when we load balance. (ECP-WarpX#1658) [mini-PR] Add particle tiling documentation (ECP-WarpX#1665) * tiling wip * do_tiling documentation * Update Docs/source/running_cpp/parameters.rst Co-authored-by: Luca Fedeli <luca.fedeli@for.unipi.it> Co-authored-by: Luca Fedeli <luca.fedeli@for.unipi.it> fix inconsistent formatting (ECP-WarpX#1666) parameter.rst fixes (ECP-WarpX#1663) requirements.txt: fix overspecification (ECP-WarpX#1668) I accidentally added the patch-level for version-compatible matching This removes the patch-level for `~=` matching for the packages that have a >=1 major version already. scipy 1.5+ (ECP-WarpX#1670) Sufficient since it was last release in december and works well. The 3.5 release series still builds wheels for Python 3.6 for Ubuntu oldstable (18.04). Remove redundant "do_qed" option in inputfile (ECP-WarpX#1667) * removed redundant do_qed option in inputfile * fixed bug remove semicolons (ECP-WarpX#1662) fixTypo for pml has particles parameter for docs (ECP-WarpX#1671) Move getCosts Always synchronize nodal points of PML MultiFabs (ECP-WarpX#1669) * Always synchronize nodal points of PML MultiFabs * Reset benchmark for CI test pml_x_psatd * Use new separate functions NodalSyncPML minor
Start a helper routine that gives performance hints after initialization of the simulation.
Example output (
amr.n_cell = 96 32 32
and 4 MPI ranks with default blocking factors of 32^3):