State Management

Parthenon manages simulation data through a hierarchy of classes designed to provide convenient state management but also high-performance in low-level, performance critical kernels. This page gives an overview of the basic classes involved in state management.

Metadata

The Metadata class provides a means of defining self-describing variables within Parthenon. It’s documentation can be found here.

StateDescriptor

The StateDescriptor class is intended to be used to inform Parthenon about the needs of an application and store relevant parameters that control application-specific behavior at runtime. The class provides several useful features and functions.

  • bool AddField(const std::string& field_name, Metadata& m) provides the means to add new (dense) variables to a Parthenon-based application with associated Metadata. This function does not allocate any storage or create any of the objects below, it simply adds the name and Metadata to a list so that those objects can be populated at the appropriate time.

  • bool AddSparsePool(...) either adds a given SparsePool or forwards the arguments to the SparsePool constructor. A SparsePool is a collection of sparse variable fields that share the same base name and Metadata, except that the shape, Vector/Tensor metadata flags, and component names can be specified per sparse id. Currently, sparse variables are allocated on all blocks just like dense variables, however, in a future upgrade, they will only be allocated on those blocks where the user explicitly allocates them or non-zero values are advected into.

  • void AddParam<T>(const std::string& key, T& value, Mutability mutability) adds a parameter (e.g., a timestep control coefficient, refinement tolerance, etc.) with name key and value value. The enum mutability can take on three values: Mutability::Immutable, Mutability::Mutable, and Mutability::Restart. Paramters that are Immutable cannot be modified. Parameters that are Mutable or Restart can be modified via the MutableParam and UpdateParam options. Parameters that are Restart will be re-read from the restart file and updated upon restart. In contrast, Mutable params not marked Restart are updated only by user code, not automatically. Note that not all parameter types can be output to HDF5 file. However, most common scalar, vector, and Kokkos view types are supported. Note also that if the value of a Param is different on different MPI ranks, this will result in undefined behaviour.

  • void AddParam<T>(const std::string& key, T& value, bool is_mutable=false) is the same as above, but adds only Immutable or Mutable params, not Restart params.

  • void UpdateParam<T>(const std::string& key, T& value)updates a parameter (e.g., a timestep control coefficient, refinement tolerance, etc.) with name key and value value. A parameter of the same type must exist.

  • const T& Param(const std::string& key) provides the getter to access parameters previously added by AddParam.

  • T *MutableParam(const std::string &key) returns a pointer to a parameter that has been marked mutable when it was added. Note this pointer is not marked const.

  • MetadataFlag GetMetadataFlag() returns a MetadataFlag that is automatically added to all fields, sparse pools, and swarms that are added to the StateDescriptor.

  • std::vector<std::string> GetVariableNames(...) provides a means of getting a list of variables that satisfy specified requirements and that are associated with the StateDescriptor. Optional arguments, in order, include a vector of variable names, a Metadata::FlagCollection to select, variables by flags, and a vector of sparse ids to allow selecting subsets of sparse fields. Selection of sparse fields by name requires providing the base name of the sparse field. Names of a sparse field tied to a specific sparse index are not supported (and will throw). The returned list is appropriate for use in adding MeshData and/or MeshBlockData objects with specified fields to the DataCollection objects in Mesh and MeshBlock. For convenience, the Mesh class also provides this function, which provides a list of variables gathered from all the package StateDescriptors.

  • void FillDerivedBlock(MeshBlockData<Real>* rc) delgates to the std::function member FillDerivedBlock if set (defaults to nullptr and therefore a no-op) that allows an application to provide a function that fills in derived quantities from independent state per MeshBlockData<Real>.

  • void FillDerivedMesh(MeshData<Real>* rc) delegates to the std::function member FillDerivedMesh if set (defaults to nullptr and therefore a no-op) that allows an application to provide a function that fills in derived quantities from independent state per MeshData<Real>.

  • Real EstimateTimestepBlock(MeshBlockData<Real>* rc) delgates to the std::function member EstimateTimestepBlock if set (defaults to nullptr and therefore a no-op) that allows an application to provide a means of computing stable/accurate timesteps for a mesh block.

  • Real EstimateTimestepMesh(MeshData<Real>* rc) delgates to the std::function member EstimateTimestepBlock if set (defaults to nullptr and therefore a no-op) that allows an application to provide a means of computing stable/accurate timesteps for a mesh block.

  • AmrTag CheckRefinement(MeshBlockData<Real>* rc) delegates to the std::function member CheckRefinementBlock if set (defaults to nullptr and therefore a no-op) that allows an application to define an application-specific refinement/de-refinement tagging function.

  • void PreStepDiagnostics(SimTime const &simtime, MeshData<Real> *rc) deletgates to the std::function member PreStepDiagnosticsMesh if set (defaults to nullptr an therefore a no-op) to print diagnostics before the time-integration advance.

  • void PostStepDiagnostics(SimTime const &simtime, MeshData<Real> *rc) deletgates to the std::function member PostStepDiagnosticsMesh if set (defaults to nullptr an therefore a no-op) to print diagnostics after the time-integration advance

  • void UserWorkBeforeLoopMesh(Mesh *, ParameterInput *pin, SimTime &tm) performs a per-package, mesh-wide calculation after (1) the mesh has been generated, (2) problem generators are called, and (3) comms are executed, but before any time evolution. This work is done both on first initialization and on restart. If you would like to avoid doing the work upon restart, you can check for the const is_restart member field of the Mesh object. It is worth making a clear distinction between UserWorkBeforeLoopMesh and ApplicationInput``s ``PostInitialization. PostInitialization is very much so tied to initialization, and will not be called upon restarts. PostInitialization is also carefully positioned after ProblemGenerator and before PreCommFillDerived (and hence communications). In practice, when additional granularity is required inbetween initialization and communication, PostInitialization may be the desired hook.

The reasoning for providing FillDerived* and EstimateTimestep* function pointers appropriate for usage with both MeshData and MeshBlockData is to allow downstream applications better control over task/kernel granularity. If, for example, the functionality needed in a package’s FillDerived* function is minimal (e.g., computing velocity from momentum and mass), better performance may be acheived by making use of the FillDerivedMesh interface. Note that applications and even individual packages can make simultaneous usage of both *Mesh and *Block functions, so long as the appropriate tasks are called as needed by the application driver.

In Parthenon, each Mesh and MeshBlock owns a Packages_t object, which is a std::map<std::string, std::shared_ptr<StateDescriptor>>. The object is intended to be populated with a StateDescriptor object per package via an Initialize function as in the advection example here. When Parthenon makes use of the Packages_t object, it iterates over all entries in the std::map. Note that it’s often useful to add a StateDescriptor to the Packages_t object for the overall application, allowing for a convenient way to define global parameters, for example.

History output

Parthenon allows packages to enroll an arbitrary number of “history” functions that are all called at the interval according to the input parameters, see output documention.

To enroll functions create a list of callback function with the appropriate reduction operation:

// List (vector) of HistoryOutputVar that will all be enrolled as output variables
parthenon::HstVar_list hst_vars = {};

// Add a callback function
hst_vars.emplace_back(parthenon::HistoryOutputVar(UserHistoryOperation::sum, MyHstFunction, "my label"));

// add callbacks for HST output identified by the `hist_param_key`
pkg->AddParam<>(parthenon::hist_param_key, hst_vars);

Here, HistoryOutputVar is a struct containing the global (over all blocks of all ranks) reduction operation, MyHstFunction is a callback function (see below), and "my label" is the string to be used as the column heading of the output file.

Currently supported reductions are

  • UserHistoryOperation::sum

  • UserHistoryOperation::min

  • UserHistoryOperation::max

which all match their respective MPI counterpart. Note, in case of volume weighting being desired (e.g., to calculate the total value in the simulation domain of some density) the volume weighting need to be done within the callback function, see the advection example.

Callback functions need to have the following signature

Real MyHstFunction(MeshData<Real> *md);

i.e., they will always work on MeshData. Note, currently history output will always be calculated for the “base” container. More specifically, the output machinery will automatically use (or create if non existent) a single “base” MeshData object containing all blocks of a rank. This simplifies the the logic for reductions over all blocks of a rank and also (generally) resuls in better performance as the number of kernel calls is reduced. However, this also implies the expectation that the “base” container holds the most recent data at the end of a timestep.

ParArrayND

This provides a light wrapper around Kokkos::View with some convenience features. It is described fully here.

Variable

The Variable class collects several associated objects that are needed to store, describe, and update simulation data. Variable is templated on type T and includes the following member data (names preceded by _ have private scope):

Member Data

Description

ParArrayND<T> data

Storage for the cell-centered associated with the object.

ParArrayND<T> flux[3]

Storage for the face-centered intercell fluxes in each direction. Only allocated for fields registered with the Metadata::Independent flag.

ParArrayND<T> coarse_s

Storage for coarse buffers need for multilevel setups.

Metadata m_

See here.

Additionally, the class overloads the () operator to provide convenient access to the data array, though this may be less efficient than operating directly on data or a reference/copy of that array.

Finally, the bool IsSet(const MetadataFlag bit) member function provides a convenient mechanism to query whether a particular Metadata flag is set for the Variable.

FaceVariable (Work in progress…)

EdgeVariable (Work in progress…)

Sparse fields

Sparse fields can be added via the StateDescriptor::AddSparsePool function. A SparsePool is a collection of sparse fields that share a common base name and metadata (see details below), but each sparse ID produces a distinct Variable. For example, a SparsePool with base name sparse and sparse IDs {3, 10, 11, 2097} will produce four Variables: sparse_3, sparse_10, sparse_11, and sparse_2097. These variables can be accessed either via their full name or the combination of base name and sparse ID. Furthermore, in a future upgrade, the sparse fields will not be allocated on all blocks but can be allocated only on specific blocks with a custom prescription on how to handle when they advect to neighboring blocks.

All the sparse field of a SparsePool share the same metadata, except for the following, which can be specified individually for each sparse ID (but they don’t have to be specified, if they are not given, they are copied from the shared metadata of the pool): - Shape - Vector/Tensor metadata flag (since that may be tied to shape) - Component labels (which is usually also tied to shape)

In particular, the associated string is shared between all sparse IDs of the same pool, so if the metadata used to create the pool has associated “foo”, then all the sparse IDs of that pool will have associated “foo”.

MeshBlockData

The MeshBlockData class provides a means of organizing and accessing simulation data. New Variables are added to a MeshBlockData container via the Add member function and accessed via various Get* functions. These Get* functions provide access to the various kinds of Variable objects described above, typically by name.

DataCollection

The DataCollection class is the highest level abstraction in Parthenon’s state management. Each MeshBlock in a simulation owns a DataCollection that through the classes just described, manages all of the simulation data. Every DataCollection is initialized with a MeshBlockData container named "base". The Get function, when invoked without arguments, returns a reference to this base MeshBlockData container which is intended to contain all of the simulation data that persists between timesteps (if applicable).

The Add(const std::string& label, MeshBlockData<T>& src) member function creates a new MeshBlockData container with the provided label. This new MeshBlockData container is populated with all of the variables in src. When a variable has the Metadata::OneCopy flag set, the variables in the new MeshBlockData container are just shallow copies from src, i.e. no new storage for data is allocated, the std::shared_ptr to the variable is just copied. For variables that do not have Metadata::OneCopy set, new storage is allocated. Once created, these new containers are accesible by calling Get with the name of the desired MeshBlockData container as an argument. NOTE: The Add function checks if a MeshBlockData container by the name label already exists in the collection, immediately returning if one is found (or throwing a std::runtime_error if the new and pre-existing containers are not equivalent). Therefore, adding a MeshBlockData container to the collection multiple times results in a single new container, with the remainder of the calls no-ops.

The overload Add(const std::string &label, MeshBlockData<T> &src, const std::vector<std::string> &names) provides the same functionality as the above Add function, but for a subset of variables provided in the vector of names. This feature allows downstream applications to allocate storage in a more targeted fashion, as might be desirable to hold source terms for particular equations, for example.

Analogously, DataCollection provides AddShallow functions that differ from Add only in that *all* included variables, even non-Metadata::OncCopy variables, are simply shallow copies. For these functions, no new storage for variables is ever allocated.

Finally, all of the functionality just described for MeshBlockData objects is also provided for MeshData objects. Adding a new MeshData object to the Mesh-level DataCollection automatically adds the corresponding MeshBlockData objects to each of the MeshBlock-level DataCollection``s.  Using this ``Mesh level functionality can be more convenient.

Two simple examples of usage of these new containers are 1) to provide storage for multistage integration schemes and 2) to provide a mechanism to allocate storage for right hand sides, deltas, etc. Both of these usages are demonstrated in the advection example that ships with Parthenon.

Note that in multistage integrator the fluxes and bvars (and their MPI communicator) of a variable are shared by default across all stages. This means that any kind of communication (most prominently flux correction and ghost zone exchange) of a given variable at a given stage should not be interleaved with any other modifications/communication of said variable as it may result in undefined behavior.