#include <MueLu_LeftoverAggregationAlgorithm_decl.hpp>

Inheritance diagram for MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >:

Private Types
typedef GO	global_size_t

typedef LO	my_size_t

Private Attributes
double	phase3AggCreation_

int	minNodesPerAggregate_

Constructors/Destructors.
	LeftoverAggregationAlgorithm ()
	Constructor. More...

virtual	~LeftoverAggregationAlgorithm ()
	Destructor. More...

Set/get methods.
void	SetMinNodesPerAggregate (int minNodesPerAggregate)

void	SetPhase3AggCreation (double phase3AggCreation)

double	GetPhase3AggCreation () const

int	GetMinNodesPerAggregate () const

Aggregation methods.
void	AggregateLeftovers (GraphBase const &graph, Aggregates &aggregates) const
	Take a partially aggregated graph and complete the aggregation. More...

Utilities
void	RootCandidates (my_size_t nVertices, ArrayView< const LO > &vertex2AggId, GraphBase const &graph, ArrayRCP< LO > &candidates, my_size_t &nCandidates, global_size_t &nCandidatesGlobal) const
	Build a list of candidate root nodes. More...

int	RemoveSmallAggs (Aggregates &aggregates, int min_size, RCP< Xpetra::Vector< double, LO, GO, NO > > &distWeights, const MueLu::CoupledAggregationCommHelper< LO, GO, NO > &myWidget) const
	Attempt to clean up aggregates that are too small. More...

Additional Inherited Members
Public Member Functions inherited from MueLu::BaseClass
virtual	~BaseClass ()
	Destructor. More...

Public Member Functions inherited from MueLu::VerboseObject
VerbLevel	GetVerbLevel () const
	Get the verbosity level. More...

void	SetVerbLevel (const VerbLevel verbLevel)
	Set the verbosity level of this object. More...

int	GetProcRankVerbose () const
	Get proc rank used for printing. Do not use this information for any other purpose. More...

int	SetProcRankVerbose (int procRank) const
	Set proc rank used for printing. More...

bool	IsPrint (MsgType type, int thisProcRankOnly=-1) const
	Find out whether we need to print out information for a specific message type. More...

Teuchos::FancyOStream &	GetOStream (MsgType type, int thisProcRankOnly=0) const
	Get an output stream for outputting the input message type. More...

Teuchos::FancyOStream &	GetBlackHole () const

	VerboseObject ()

virtual	~VerboseObject ()
	Destructor. More...

Public Member Functions inherited from Teuchos::VerboseObject< VerboseObject >
	VerboseObject (const EVerbosityLevel verbLevel=VERB_DEFAULT, const RCP< FancyOStream > &oStream=Teuchos::null)

virtual const VerboseObject &	setVerbLevel (const EVerbosityLevel verbLevel) const

virtual const VerboseObject &	setOverridingVerbLevel (const EVerbosityLevel verbLevel) const

virtual EVerbosityLevel	getVerbLevel () const

TEUCHOSPARAMETERLIST_LIB_DLL_EXPORT RCP< const ParameterList >	getValidVerboseObjectSublist ()

TEUCHOSPARAMETERLIST_LIB_DLL_EXPORT void	setupVerboseObjectSublist (ParameterList *paramList)

TEUCHOSPARAMETERLIST_LIB_DLL_EXPORT void	readVerboseObjectSublist (ParameterList paramList, RCP< FancyOStream > oStream, EVerbosityLevel *verbLevel)

void	readVerboseObjectSublist (ParameterList paramList, VerboseObject< ObjectType > verboseObject)

Public Member Functions inherited from Teuchos::VerboseObjectBase
virtual	~VerboseObjectBase ()

	VerboseObjectBase (const RCP< FancyOStream > &oStream=Teuchos::null)

virtual const VerboseObjectBase &	setOStream (const RCP< FancyOStream > &oStream) const

virtual const VerboseObjectBase &	setOverridingOStream (const RCP< FancyOStream > &oStream) const

virtual VerboseObjectBase &	setLinePrefix (const std::string &linePrefix)

virtual RCP< FancyOStream >	getOStream () const

virtual RCP< FancyOStream >	getOverridingOStream () const

virtual std::string	getLinePrefix () const

virtual OSTab	getOSTab (const int tabs=1, const std::string &linePrefix="") const

Public Member Functions inherited from MueLu::Describable
virtual	~Describable ()
	Destructor. More...

virtual std::string	ShortClassName () const
	Return the class name of the object, without template parameters and without namespace. More...

virtual void	describe (Teuchos::FancyOStream &out_arg, const VerbLevel verbLevel=Default) const

virtual std::string	description () const
	Return a simple one-line description of this object. More...

void	describe (Teuchos::FancyOStream &out, const Teuchos::EVerbosityLevel verbLevel=Teuchos::Describable::verbLevel_default) const
	Print the object with some verbosity level to an FancyOStream object. More...

Public Member Functions inherited from Teuchos::Describable
void	describe (std::ostream &out, const EVerbosityLevel verbLevel=verbLevel_default) const

	LabeledObject ()

virtual	~LabeledObject ()

virtual void	setObjectLabel (const std::string &objectLabel)

virtual std::string	getObjectLabel () const

DescribableStreamManipulatorState	describe (const Describable &describable, const EVerbosityLevel verbLevel=Describable::verbLevel_default)

std::ostream &	operator<< (std::ostream &os, const DescribableStreamManipulatorState &d)

Static Public Member Functions inherited from MueLu::VerboseObject
static void	SetMueLuOStream (const Teuchos::RCP< Teuchos::FancyOStream > &mueluOStream)

static Teuchos::RCP < Teuchos::FancyOStream >	GetMueLuOStream ()

static void	SetDefaultVerbLevel (const VerbLevel defaultVerbLevel)
	Set the default (global) verbosity level. More...

static VerbLevel	GetDefaultVerbLevel ()
	Get the default (global) verbosity level. More...

Static Public Member Functions inherited from Teuchos::VerboseObject< VerboseObject >
static void	setDefaultVerbLevel (const EVerbosityLevel defaultVerbLevel)

static EVerbosityLevel	getDefaultVerbLevel ()

Static Public Member Functions inherited from Teuchos::VerboseObjectBase
static void	setDefaultOStream (const RCP< FancyOStream > &defaultOStream)

static RCP< FancyOStream >	getDefaultOStream ()

Static Public Attributes inherited from Teuchos::Describable
static const EVerbosityLevel	verbLevel_default

Protected Member Functions inherited from Teuchos::VerboseObject< VerboseObject >
void	initializeVerboseObject (const EVerbosityLevel verbLevel=VERB_DEFAULT, const RCP< FancyOStream > &oStream=Teuchos::null)

Protected Member Functions inherited from Teuchos::VerboseObjectBase
void	initializeVerboseObjectBase (const RCP< FancyOStream > &oStream=Teuchos::null)

virtual void	informUpdatedVerbosityState () const

Detailed Description

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>
class MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >

Definition at line 64 of file MueLu_LeftoverAggregationAlgorithm_decl.hpp.

Member Typedef Documentation

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>

typedef GO MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::global_size_t

private

Definition at line 68 of file MueLu_LeftoverAggregationAlgorithm_decl.hpp.

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>

typedef LO MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::my_size_t

private

Definition at line 69 of file MueLu_LeftoverAggregationAlgorithm_decl.hpp.

Constructor & Destructor Documentation

template<class LocalOrdinal , class GlobalOrdinal , class Node >

MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::LeftoverAggregationAlgorithm ( )

Constructor.

Definition at line 65 of file MueLu_LeftoverAggregationAlgorithm_def.hpp.

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>

virtual MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::~LeftoverAggregationAlgorithm ( )

inlinevirtual

Destructor.

Definition at line 80 of file MueLu_LeftoverAggregationAlgorithm_decl.hpp.

Member Function Documentation

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>

void MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::SetMinNodesPerAggregate ( int minNodesPerAggregate )

inline

Definition at line 87 of file MueLu_LeftoverAggregationAlgorithm_decl.hpp.

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>

void MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::SetPhase3AggCreation ( double phase3AggCreation )

inline

Definition at line 88 of file MueLu_LeftoverAggregationAlgorithm_decl.hpp.

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>

double MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::GetPhase3AggCreation ( ) const

inline

Definition at line 90 of file MueLu_LeftoverAggregationAlgorithm_decl.hpp.

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>

int MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::GetMinNodesPerAggregate ( ) const

inline

Definition at line 91 of file MueLu_LeftoverAggregationAlgorithm_decl.hpp.

template<class LocalOrdinal , class GlobalOrdinal , class Node >

void MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::AggregateLeftovers	(	GraphBase const &	graph,
		Aggregates &	aggregates
	)		const

Take a partially aggregated graph and complete the aggregation.

This is typically needed to take care of vertices that are left over after creating a bunch of ideal aggregates (root plus immediate neighbors).

On input, the structure Aggregates describes already aggregated vertices. The field procWinners[] indicates the processor owning the aggregate to which a vertex is "definitively" assigned. If on entry procWinners[i] == MUELU_UNASSIGNED, ArbitrateAndCommunicate() will arbitrate and decide which processor's aggregate really has the vertex. If only one processor claims ownership (as in the Uncoupled case), no real arbitration is needed. Otherwise, random arbitration is done.

This cleanup has many phases:

Phase 1b: Invoke ArbitrateAndCommunicate() to ensure that all processors have the same view of aggregated vertices (e.g., to which aggregate they have been assigned and which processor owns that aggregate).
Phase 2: Check for vertices (local or nonlocal) which are Adjacent to root nodes. Tentatively assign these to the aggregate associated with the root. Arbitrate any cases where several processors claim the same vertex for one of their aggregates via ArbitrateAndCommunicate().

Phase 3: Try to create new aggregates if it looks like there are root node candidates which have many unaggregated neighbors. This decision to make a new aggregate is based on only local information. However, the new aggregate will be tentatively assigned any unaggregated ghost vertices. Arbitration is again done by ArbitrateAndCommunicate() where local vertices use a weight[] = 2 and ghost vertices have weight[] = 1. The basic idea is that after arbitration, each aggregate is guaranteed to keep all local vertices assigned in this phase. Thus, by basing the aggregation creation logic on local information, we are guaranteed to have a sufficiently large aggregation. The only local vertices that might be assigned to another processor's aggregates are unclaimed during this phase of the aggregation.

Phase 4: EXPERIMENTAL
Phase 5: Sweep new points into existing aggregates. Each processor tries to assign any (whether it is a ghost or local) unaggregated vertex that it has to an aggregate that it owns. In other words, processor p attempts to assign vertex \(v\) to aggregate \(y\) where \(y\) is owned by p but \(v\) may be a ghost vertex (and so really assigned to another processor). Deciding which aggregate a vertex is assigned to is done by scoring. Roughly, we want
- larger scores for \(y\) if \(v\) is close (graph distance) to \(y\)'s root.
- larger scores for \(y\) if \(v\) has direct connections to several different vertices already assigned to \(y\).
- lower scores for \(y\) if several vertices have already been swept into \(y\) during this phase.

Some care must be taken for vertices that are shared (either local vertices that are sent to other processors or ghost vertices) in that each processor sharing the vertex will attempt to assign it to different aggregates. ArbitrateAndCommunicate() is again used for arbitration with the score being given as the weight.

The main tricky thing occurs when \(v\) is tentatively added to \(y\). When assigning \(v'\) to \(y\), the assumed connection with \(v\) should not be the sole basis of this decision if there is some chance that \(v\) might be lost in arbitration. This could actually lead to \(v'\) being disconnected from the rest of the aggregate. This is by building a list of shared ids and checking that there is at least one vertex in the scoring that is either not shared or has already been definitively assigned to this processor's aggregate (i.e. have been assigned to a local aggregate and have been through arbitration).

Scoring is done by first giving a mark to vertices that have been already been assigned to aggregates. This mark essentially reflects the distance of this point to the root. Specifically,

\[ mark(v) \leftarrow MUELU\_DISTONE\_VERTEX\_WEIGHT \]

if \(v\) was assigned to an aggregate prior to this phase,

\[ mark(v) \leftarrow max(mark(v_k))/2 \]

otherwise, where \(max(mark(v_k))\) considers all vertices definitively assigned to \(y\) that have direct connections to \(v\).

Finally,

\[ score(\tilde{v},y) \leftarrow \Sigma(mark(\hat{v}_k)) - AggregateIncrementPenalty \]

where \(\tilde{v}\) is an unaggregated vertex being considered for assignment in aggregate \(y\) and \(hat{v}_k\) are all vertices in \(y\) with a direct connection to \(\tilde{v}\). AggregateIncrementPenalty is equal to

\[ \max (\mbox{INCR_SCALING}*NNewVtx, \Sigma(mark(\hat{v}_k))*(1-\mbox{MUELU_PENALTYFACTOR})) \]

where \( NNewVtx \) is the number of phase 5 vertices already assigned to \(y\).

One last wrinkle, is that we have wrapped the whole scoring/assigning of vertices inside a big loop that looks something like

for ( Threshold = big; Threshold >= 0; Reduce(Threshold));

New vertices are swept into aggregates only if their best score is >= a Threshold. This encourages only very good vertices to be assigned first followed by somewhat less well connected ones in later iterations of the loop. It also helps minimize the number of exclusions that would occur to address the issue mentioned above where we don't want to make assignment decisions based on connections to vertices that might be later lost in arbitration.

Phase 6: Aggregate remaining vertices and avoid small aggregates (e.g., singletons) at all costs. Typically, most everything should be aggregated by Phases 1-5. One way that we could still have unaggregated vertices is if processor p was never assigned a root node (perhaps because the number of local vertices on p is less than minNodesPerAggregate) and additionally p has local ids which are not shared with any other processors (so that no other processor's aggregate can claim these vertices).

Phase 6 looks at the first unassigned vertex and all of its local unassigned neighbors and makes a new aggregate. If this aggregate has at least minNodesPerAggregate vertices, we continue this process of creating new aggregates by examining other unassigned vertices. If the new aggregate is too small, we try add the next unassigned vertex and its neighbors to the same newly created aggregate. Once again, we check the size of this new aggregate to decide whether other unassigned vertices should be added to this aggregate or used to create a new aggregate. If the last newly created aggregate (of course there may be just one newly created aggregate) is too small, we then see if there is at least one other aggregate that this processor owns. If so, we merge the small newly created aggregate with aggregate 0. If not, we accept the fact that a small aggregate has been created.

One final note about the use of ArbitrateAndCommunicate(). No arbitration occurs (which means the procWinner[] is not set as well) for a global shared id if and only if all weights on all processors corresponding to this id is zero. Thus, the general idea is that any global id where we want arbitration to occur should have at least one associated weight on one processor which is nonzero. Any global id where we don't want any arbitration should have all weights set to 0.

Note: procWinners is also set to MyPid() by ArbitrateAndCommunicate() for any nonshared gid's with a nonzero weight.

Definition at line 71 of file MueLu_LeftoverAggregationAlgorithm_def.hpp.

template<class LocalOrdinal , class GlobalOrdinal , class Node >

void MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::RootCandidates	(	my_size_t	nVertices,
		ArrayView< const LO > &	vertex2AggId,
		GraphBase const &	graph,
		ArrayRCP< LO > &	candidates,
		my_size_t &	nCandidates,
		global_size_t &	nCandidatesGlobal
	)		const

Build a list of candidate root nodes.

Candidates are vertices not adjacent to already aggregated vertices.

Definition at line 675 of file MueLu_LeftoverAggregationAlgorithm_def.hpp.

template<class LocalOrdinal , class GlobalOrdinal , class Node >

int MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::RemoveSmallAggs	(	Aggregates &	aggregates,
		int	min_size,
		RCP< Xpetra::Vector< double, LO, GO, NO > > &	distWeights,
		const MueLu::CoupledAggregationCommHelper< LO, GO, NO > &	myWidget
	)		const

Attempt to clean up aggregates that are too small.

Definition at line 702 of file MueLu_LeftoverAggregationAlgorithm_def.hpp.

Member Data Documentation

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>

double MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::phase3AggCreation_

private

Definition at line 311 of file MueLu_LeftoverAggregationAlgorithm_decl.hpp.

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>

int MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >::minNodesPerAggregate_

private

Definition at line 312 of file MueLu_LeftoverAggregationAlgorithm_decl.hpp.

The documentation for this class was generated from the following files:

Private Types

Private Attributes

Constructors/Destructors.

Set/get methods.

Aggregation methods.

Utilities

Additional Inherited Members

Detailed Description

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType> class MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >

Member Typedef Documentation

Constructor & Destructor Documentation

Member Function Documentation

Member Data Documentation

template<class LocalOrdinal = int, class GlobalOrdinal = LocalOrdinal, class Node = KokkosClassic::DefaultNode::DefaultNodeType>
class MueLu::LeftoverAggregationAlgorithm< LocalOrdinal, GlobalOrdinal, Node >