class for distributing obs across multiple process elements More...
#include <Distribution.h>
Public Member Functions | |
| Distribution (const eckit::mpi::Comm &Comm) | |
| virtual | ~Distribution () |
| virtual bool | isIdentity () const |
| Returns true if the distribution assigns all records to all PEs, false otherwise. More... | |
| virtual bool | isNonoverlapping () const |
| Returns true if the distribution does not assign any record to more than one PE, false otherwise. More... | |
| virtual void | assignRecord (const std::size_t RecNum, const std::size_t LocNum, const eckit::geometry::Point2 &point) |
If the record RecNum has not yet been assigned to a PE, assigns it to the appropriate PE. More... | |
| virtual bool | isMyRecord (std::size_t RecNum) const =0 |
Returns true if record RecNum has been assigned to the calling PE during a previous call to assignRecord(). More... | |
| virtual void | computePatchLocs () |
| If necessary, identifies locations of "patch obs", i.e. More... | |
| virtual void | patchObs (std::vector< bool > &isPatchObs) const =0 |
| Sets each element of the provided vector to true if the corresponding location is a "patch obs", i.e. More... | |
| virtual void | min (int &x) const =0 |
| Calculates the global minimum (over all locations on all PEs) of a location-dependent quantity. More... | |
| virtual void | min (std::size_t &x) const =0 |
| virtual void | min (float &x) const =0 |
| virtual void | min (double &x) const =0 |
| virtual void | min (std::vector< int > &x) const =0 |
| Calculates the global minima (over all locations on all PEs) of multiple location-dependent quantities. More... | |
| virtual void | min (std::vector< std::size_t > &x) const =0 |
| virtual void | min (std::vector< float > &x) const =0 |
| virtual void | min (std::vector< double > &x) const =0 |
| virtual void | max (int &x) const =0 |
| Calculates the global maximum (over all locations on all PEs) of a location-dependent quantity. More... | |
| virtual void | max (std::size_t &x) const =0 |
| virtual void | max (float &x) const =0 |
| virtual void | max (double &x) const =0 |
| virtual void | max (std::vector< int > &x) const =0 |
| Calculates the global maxima (over all locations on all PEs) of multiple location-dependent quantities. More... | |
| virtual void | max (std::vector< std::size_t > &x) const =0 |
| virtual void | max (std::vector< float > &x) const =0 |
| virtual void | max (std::vector< double > &x) const =0 |
| template<typename T > | |
| std::unique_ptr< Accumulator< T > > | createAccumulator () const |
| Create an object that can be used to calculate the sum of a location-dependent quantity over locations held on all PEs, each taken into account only once even if it's held on multiple PEs. More... | |
| template<typename T > | |
| std::unique_ptr< Accumulator< std::vector< T > > > | createAccumulator (std::size_t n) const |
Create an object that can be used to calculate the sums of n location-dependent quantities over locations held on all PEs, each taken into account only once even if it's held on multiple PEs. More... | |
| virtual void | allGatherv (std::vector< size_t > &x) const =0 |
| Gather observation data from all processes and deliver the combined data to all processes. More... | |
| virtual void | allGatherv (std::vector< int > &x) const =0 |
| virtual void | allGatherv (std::vector< float > &x) const =0 |
| virtual void | allGatherv (std::vector< double > &x) const =0 |
| virtual void | allGatherv (std::vector< util::DateTime > &x) const =0 |
| virtual void | allGatherv (std::vector< std::string > &x) const =0 |
| virtual size_t | globalUniqueConsecutiveLocationIndex (size_t loc) const =0 |
| Map the index of a location held on the calling process to the index of the corresponding element of any vector produced by allGatherv(). More... | |
| virtual std::string | name () const =0 |
| size_t | rank () const |
| Accessor to MPI rank. More... | |
Protected Attributes | |
| const eckit::mpi::Comm & | comm_ |
| Local MPI communicator. More... | |
Private Member Functions | |
| virtual std::unique_ptr< Accumulator< int > > | createAccumulatorImpl (int init) const =0 |
| Create an object that can be used to calculate the sum of a location-dependent quantity over locations held on all PEs, each taken into account only once even if it's held on multiple PEs. More... | |
| virtual std::unique_ptr< Accumulator< std::size_t > > | createAccumulatorImpl (std::size_t init) const =0 |
| virtual std::unique_ptr< Accumulator< float > > | createAccumulatorImpl (float init) const =0 |
| virtual std::unique_ptr< Accumulator< double > > | createAccumulatorImpl (double init) const =0 |
| virtual std::unique_ptr< Accumulator< std::vector< int > > > | createAccumulatorImpl (const std::vector< int > &init) const =0 |
| Create an object that can be used to calculate the sums of multiple location-dependent quantities over locations held on all PEs, each taken into account only once even if it's held on multiple PEs. More... | |
| virtual std::unique_ptr< Accumulator< std::vector< std::size_t > > > | createAccumulatorImpl (const std::vector< std::size_t > &init) const =0 |
| virtual std::unique_ptr< Accumulator< std::vector< float > > > | createAccumulatorImpl (const std::vector< float > &init) const =0 |
| virtual std::unique_ptr< Accumulator< std::vector< double > > > | createAccumulatorImpl (const std::vector< double > &init) const =0 |
Related Functions | |
(Note that these are not member functions.) | |
| double | dotProduct (const Distribution &dist, std::size_t numVariables, const std::vector< double > &v1, const std::vector< double > &v2) |
| Computes the dot product between two vectors of obs distributed across MPI ranks. More... | |
| std::size_t | globalNumNonMissingObs (const Distribution &dist, size_t numVariables, const std::vector< double > &v) |
| Counts unique non-missing observations in a vector. More... | |
Detailed Description
class for distributing obs across multiple process elements
This Distribution class is a base class where various subclasses of this class define different methods for distributing obs.
The subclasses of this base class need to implement all pure virtual functions. The client will use the isMyRecord method to determine what records to keep when reading in observations.
assignRecord() must be called before isMyRecord() can be called. computePatchLocs() should be called when all records have been assigned.
for distributions with observations duplicated on multiple PEs (currently Inefficient and Halo) the following terminology/logic is used: the set of obs on each PE is called halo obs. (as in within a halo of some location) obs in the halo set can be duplicated across multiple PEs the subset of halo obs. (patch obs.) "belong" only to this PE and are used to compute reduce operations without duplication. patch obs form complete, non-overlapping partition of the global set of obs.
- See also
- Functions defined in DistributionUtils.h.
Definition at line 56 of file src/distribution/Distribution.h.
Constructor & Destructor Documentation
◆ Distribution()
|
explicit |
Definition at line 19 of file Distribution.cc.
◆ ~Distribution()
|
virtual |
Definition at line 25 of file Distribution.cc.
Member Function Documentation
◆ allGatherv() [1/6]
|
pure virtual |
◆ allGatherv() [2/6]
|
pure virtual |
◆ allGatherv() [3/6]
|
pure virtual |
◆ allGatherv() [4/6]
|
pure virtual |
Gather observation data from all processes and deliver the combined data to all processes.
- Parameters
-
[in,out] x On input: a vector whose ith element is associated with the ith observation held by the calling process. On output: a concatenation of the vectors xpassed by all calling processes, with duplicates removed (i.e. if any observations are duplicated across multiple processes, the elements ofxcorresponding to these data are included only once).
Implemented in ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::NonoverlappingDistribution, ioda::InefficientDistribution, and ioda::Halo.
◆ allGatherv() [5/6]
|
pure virtual |
◆ allGatherv() [6/6]
|
pure virtual |
◆ assignRecord()
|
inlinevirtual |
If the record RecNum has not yet been assigned to a PE, assigns it to the appropriate PE.
Informs the distribution that location LocNum belongs to this record.
- Parameters
-
RecNum Record containing the location LocNum.LocNum (Global) location index. point Latitude and longitude of this location.
Reimplemented in ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::NonoverlappingDistribution, ioda::Halo, and ioda::AtlasDistribution.
Definition at line 80 of file src/distribution/Distribution.h.
◆ computePatchLocs()
|
inlinevirtual |
If necessary, identifies locations of "patch obs", i.e.
locations belonging to records owned by this PE.
This function must be called when all records have been assigned, and in particular before any calls to the createAccumulator() and globalUniqueConsecutiveLocationIndex() member functions or the global functions and dotProduct(), globalNumNonMissingObs().
Reimplemented in ioda::ReplicaOfGeneralDistribution, ioda::NonoverlappingDistribution, and ioda::Halo.
Definition at line 100 of file src/distribution/Distribution.h.
◆ createAccumulator() [1/2]
|
inline |
Create an object that can be used to calculate the sum of a location-dependent quantity over locations held on all PEs, each taken into account only once even if it's held on multiple PEs.
- Template Parameters
-
T Must be either int,size_t,floatordouble.
- Parameters
-
init Used only to select the right overload – the value is ignored.
Definition at line 175 of file src/distribution/Distribution.h.
◆ createAccumulator() [2/2]
|
inline |
Create an object that can be used to calculate the sums of n location-dependent quantities over locations held on all PEs, each taken into account only once even if it's held on multiple PEs.
- Template Parameters
-
T Must be either int,size_t,floatordouble.
- Parameters
-
n The number of quantities to be summed up.
Definition at line 193 of file src/distribution/Distribution.h.
◆ createAccumulatorImpl() [1/8]
|
privatepure virtual |
◆ createAccumulatorImpl() [2/8]
|
privatepure virtual |
◆ createAccumulatorImpl() [3/8]
|
privatepure virtual |
Create an object that can be used to calculate the sums of multiple location-dependent quantities over locations held on all PEs, each taken into account only once even if it's held on multiple PEs.
- Parameters
-
init A vector of as many elements as there are sums to calculate. The values of these elements are ignored.
Implemented in ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::NonoverlappingDistribution, ioda::InefficientDistribution, and ioda::Halo.
◆ createAccumulatorImpl() [4/8]
|
privatepure virtual |
◆ createAccumulatorImpl() [5/8]
|
privatepure virtual |
◆ createAccumulatorImpl() [6/8]
|
privatepure virtual |
◆ createAccumulatorImpl() [7/8]
|
privatepure virtual |
Create an object that can be used to calculate the sum of a location-dependent quantity over locations held on all PEs, each taken into account only once even if it's held on multiple PEs.
- Parameters
-
init Used only to select the right overload – the value is ignored.
Implemented in ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::NonoverlappingDistribution, ioda::InefficientDistribution, and ioda::Halo.
◆ createAccumulatorImpl() [8/8]
|
privatepure virtual |
◆ globalUniqueConsecutiveLocationIndex()
|
pure virtual |
Map the index of a location held on the calling process to the index of the corresponding element of any vector produced by allGatherv().
Implemented in ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::NonoverlappingDistribution, ioda::InefficientDistribution, and ioda::Halo.
◆ isIdentity()
|
inlinevirtual |
Returns true if the distribution assigns all records to all PEs, false otherwise.
Reimplemented in ioda::InefficientDistribution.
Definition at line 64 of file src/distribution/Distribution.h.
◆ isMyRecord()
|
pure virtual |
Returns true if record RecNum has been assigned to the calling PE during a previous call to assignRecord().
Clients can use this function to determine which records to keep when reading in observations.
Implemented in ioda::RoundRobin, ioda::ReplicaOfNonoverlappingDistribution, ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::InefficientDistribution, ioda::Halo, and ioda::AtlasDistribution.
◆ isNonoverlapping()
|
inlinevirtual |
Returns true if the distribution does not assign any record to more than one PE, false otherwise.
Reimplemented in ioda::NonoverlappingDistribution.
Definition at line 70 of file src/distribution/Distribution.h.
◆ max() [1/8]
|
pure virtual |
◆ max() [2/8]
|
pure virtual |
◆ max() [3/8]
|
pure virtual |
Calculates the global maximum (over all locations on all PEs) of a location-dependent quantity.
- Parameters
-
[in,out] x On input, the local maximum (over all locations on the current PE) of a location-dependent quantity. On output, its global maximum.
Implemented in ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::NonoverlappingDistribution, ioda::InefficientDistribution, and ioda::Halo.
◆ max() [4/8]
|
pure virtual |
◆ max() [5/8]
|
pure virtual |
◆ max() [6/8]
|
pure virtual |
◆ max() [7/8]
|
pure virtual |
Calculates the global maxima (over all locations on all PEs) of multiple location-dependent quantities.
- Parameters
-
[in,out] x On input, each element of xshould be the local maximum (over all locations on the current PE) of a location-dependent quantity. On output, that element will be set to the global maximum of that quantity.
Implemented in ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::NonoverlappingDistribution, ioda::InefficientDistribution, and ioda::Halo.
◆ max() [8/8]
|
pure virtual |
◆ min() [1/8]
|
pure virtual |
◆ min() [2/8]
|
pure virtual |
◆ min() [3/8]
|
pure virtual |
Calculates the global minimum (over all locations on all PEs) of a location-dependent quantity.
- Parameters
-
[in,out] x On input, the local minimum (over all locations on the current PE) of a location-dependent quantity. On output, its global minimum.
Implemented in ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::NonoverlappingDistribution, ioda::InefficientDistribution, and ioda::Halo.
◆ min() [4/8]
|
pure virtual |
◆ min() [5/8]
|
pure virtual |
◆ min() [6/8]
|
pure virtual |
◆ min() [7/8]
|
pure virtual |
Calculates the global minima (over all locations on all PEs) of multiple location-dependent quantities.
- Parameters
-
[in,out] x On input, each element of xshould be the local minimum (over all locations on the current PE) of a location-dependent quantity. On output, that element will be set to the global minimum of that quantity.
Implemented in ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::NonoverlappingDistribution, ioda::InefficientDistribution, and ioda::Halo.
◆ min() [8/8]
|
pure virtual |
◆ name()
|
pure virtual |
Implemented in ioda::RoundRobin, ioda::ReplicaOfNonoverlappingDistribution, ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::InefficientDistribution, ioda::Halo, and ioda::AtlasDistribution.
◆ patchObs()
|
pure virtual |
Sets each element of the provided vector to true if the corresponding location is a "patch obs", i.e.
it belongs to a record owned by this PE, and to false otherwise.
- Parameters
-
isPatchObs Preallocated vector of as many elements as there are locations on this PE.
Implemented in ioda::NonoverlappingDistribution, ioda::ReplicaOfGeneralDistribution, ioda::PairOfDistributions, ioda::InefficientDistribution, and ioda::Halo.
◆ rank()
|
inline |
Accessor to MPI rank.
Definition at line 228 of file src/distribution/Distribution.h.
Friends And Related Function Documentation
◆ dotProduct()
|
related |
Computes the dot product between two vectors of obs distributed across MPI ranks.
- Parameters
-
distribution Distribution used to partition observations across MPI ranks. numVariables Number of variables whose observations are stored in v1andv2.v1,v2 Vectors of observations. Observations of individual variables should be interleaved, i.e. the observation of variable ivarat locationilocin the halo of the calling MPI rank should be stored in element(iloc * numVariables + ivar)of each vector.
- Returns
- The dot product of the two vectors, with observations taken at locations belonging to the halos of multiple MPI ranks counted only once and any missing values treated as if they were zeros.
Definition at line 68 of file DistributionUtils.cc.
◆ globalNumNonMissingObs()
|
related |
Counts unique non-missing observations in a vector.
- Parameters
-
distribution Distribution used to partition observations across MPI ranks. numVariables Number of variables whose observations are stored in v.v Vector of observations. Observations of individual variables should be interleaved, i.e. the observation of variable ivarat locationilocin the halo of the calling MPI rank should be stored atv[iloc * numVariables + ivar].
- Returns
- The number of unique observations on all MPI ranks set to something else than the missing value indicator. "Unique" means that observations taken at locations belonging to the halos of multiple MPI ranks are counted only once.
Member Data Documentation
◆ comm_
|
protected |
Local MPI communicator.
Definition at line 268 of file src/distribution/Distribution.h.
The documentation for this class was generated from the following files:
