Supervisor of parallelism using MPI and Threads. More...

Inheritance diagram for Arcane::MessagePassing::HybridParallelSuperMng:

Collaboration diagram for Arcane::MessagePassing::HybridParallelSuperMng:

Public Member Functions
	HybridParallelSuperMng (const ServiceBuildInfo &sbi)
void	initialize () override
	Initializes the instance.
void	build () override
	Constructs the instance members.
IApplication *	application () const override
	Returns the main manager.
IThreadMng *	threadMng () const override
	Thread manager.
bool	isParallel () const override
	Returns true if the execution is parallel.
Int32	commRank () const override
	Returns the process number (between 0 and nbProcess()-1).
Int32	commSize () const override
	Returns the total number of processes used.
Int32	traceRank () const override
	Rank of this instance for traces.
void *	getMPICommunicator () override
	Address of the MPI communicator associated with this manager.
MP::Communicator	communicator () const override
	MPI communicator associated with this manager.
Ref< IParallelMng >	internalCreateWorldParallelMng (Int32 local_rank) override
	Creates a parallelism manager for all allocated cores.
void	tryAbort () override
	Attempts to abort.
bool	isMasterIO () const override
	Returns true if the instance is a master I/O manager.
Integer	masterIORank () const override
	Rank of the instance managing input/output (for which isMasterIO() is true).
Integer	nbLocalSubDomain () override
	Number of subdomains to create locally.
void	barrier () override
	Parallelism manager for all allocated resources.
Public Member Functions inherited from Arcane::ParallelSuperMngDispatcher
virtual void	allGather (ByteConstArrayView send_buf, ByteArrayView recv_buf)
virtual void	allGather (Int32ConstArrayView send_buf, Int32ArrayView recv_buf)
virtual void	allGather (Int64ConstArrayView send_buf, Int64ArrayView recv_buf)
virtual void	allGather (RealConstArrayView send_buf, RealArrayView recv_buf)
virtual Int32	reduce (eReduceType rt, Int32 v)
virtual Int64	reduce (eReduceType rt, Int64 v)
virtual Real	reduce (eReduceType rt, Real v)
virtual void	reduce (eReduceType rt, Int32ArrayView v)
virtual void	reduce (eReduceType rt, Int64ArrayView v)
virtual void	reduce (eReduceType rt, RealArrayView v)
void	broadcast (ByteArrayView send_buf, Integer id) override
	Sends an array of values to all processes This operation synchronizes the value array send_buf across all processes. The array used is that of the process whose identifier (processId()) is process_id. All processes must call this method with the same parameter process_id and have a send_buf array containing the same number of elements.
void	broadcast (Int32ArrayView send_buf, Integer id) override
void	broadcast (Int64ArrayView send_buf, Integer id) override
void	broadcast (RealArrayView send_buf, Integer id) override
virtual void	send (ByteConstArrayView values, Integer id)
virtual void	send (Int32ConstArrayView values, Integer id)
virtual void	send (Int64ConstArrayView values, Integer id)
virtual void	send (RealConstArrayView values, Integer id)
virtual void	recv (ByteArrayView values, Integer id)
virtual void	recv (Int32ArrayView values, Integer id)
virtual void	recv (Int64ArrayView values, Integer id)
virtual void	recv (RealArrayView values, Integer id)
virtual Request	send (ByteConstArrayView values, Integer id, bool is_blocked)
virtual Request	send (Int32ConstArrayView values, Integer id, bool is_blocked)
virtual Request	send (Int64ConstArrayView values, Integer id, bool is_blocked)
virtual Request	send (RealConstArrayView values, Integer id, bool is_blocked)
virtual Request	recv (ByteArrayView values, Integer id, bool is_blocked)
virtual Request	recv (Int32ArrayView values, Integer id, bool is_blocked)
virtual Request	recv (Int64ArrayView values, Integer id, bool is_blocked)
virtual Request	recv (RealArrayView values, Integer id, bool is_blocked)
virtual void	sendRecv (ByteConstArrayView send_buf, ByteArrayView recv_buf, Integer id)
virtual void	sendRecv (Int32ConstArrayView send_buf, Int32ArrayView recv_buf, Integer id)
virtual void	sendRecv (Int64ConstArrayView send_buf, Int64ArrayView recv_buf, Integer id)
virtual void	sendRecv (RealConstArrayView send_buf, RealArrayView recv_buf, Integer id)
virtual void	allToAll (ByteConstArrayView send_buf, ByteArrayView recv_buf, Integer count)
virtual void	allToAll (Int32ConstArrayView send_buf, Int32ArrayView recv_buf, Integer count)
virtual void	allToAll (Int64ConstArrayView send_buf, Int64ArrayView recv_buf, Integer count)
virtual void	allToAll (RealConstArrayView send_buf, RealArrayView recv_buf, Integer count)
virtual Int32	scan (eReduceType rt, Int32 v)
virtual Int64	scan (eReduceType rt, Int64 v)
virtual Real	scan (eReduceType rt, Real v)
virtual void	scan (eReduceType rt, Int32ArrayView v)
virtual void	scan (eReduceType rt, Int64ArrayView v)
virtual void	scan (eReduceType rt, RealArrayView v)
Public Member Functions inherited from Arcane::IParallelSuperMng
virtual	~IParallelSuperMng ()
	Frees resources.

Public Attributes
HybridParallelMngContainer *	m_container = nullptr
Ref< IParallelMngContainerFactory >	m_builder_factory
Ref< IParallelMngContainer >	m_main_builder
IApplication *	m_application
	Main manager.
Parallel::IStat *	m_stat
Int32	m_mpi_comm_rank = -1
	Statistics.
Int32	m_mpi_comm_size = -1
	Number of processors.
MPI_Comm	m_mpi_communicator
MP::Communicator	m_communicator
MP::Communicator	m_machine_communicator
Int32	m_local_nb_rank = A_NULL_RANK
MpiLock *	m_mpi_lock
MpiAdapter *	m_mpi_adapter
MpiDatatypeList *	m_datatype_list
MpiErrorHandler	m_error_handler

Additional Inherited Members
Public Types inherited from Arcane::IParallelSuperMng
typedef Parallel::Request	Request
typedef Parallel::eReduceType	eReduceType
Protected Member Functions inherited from Arcane::ParallelSuperMngDispatcher
void	_setDispatchers (IParallelDispatchT< Byte > b, IParallelDispatchT< Int32 > i32, IParallelDispatchT< Int64 > i64, IParallelDispatchT< Real > r)
void	_finalize ()

Detailed Description

Supervisor of parallelism using MPI and Threads.

Definition at line 302 of file HybridParallelSuperMng.cc.

Constructor & Destructor Documentation

◆ HybridParallelSuperMng()

Arcane::MessagePassing::HybridParallelSuperMng::HybridParallelSuperMng ( const ServiceBuildInfo & sbi )

explicit

Definition at line 351 of file HybridParallelSuperMng.cc.

◆ ~HybridParallelSuperMng()

Arcane::MessagePassing::HybridParallelSuperMng::~HybridParallelSuperMng ( )

override

Definition at line 368 of file HybridParallelSuperMng.cc.

Member Function Documentation

◆ application()

IApplication * Arcane::MessagePassing::HybridParallelSuperMng::application ( ) const

inlineoverridevirtual

Returns the main manager.

Implements Arcane::IParallelSuperMng.

Definition at line 313 of file HybridParallelSuperMng.cc.

References m_application.

◆ barrier()

void Arcane::MessagePassing::HybridParallelSuperMng::barrier ( )

overridevirtual

Parallelism manager for all allocated resources.

Performs a barrier

Implements Arcane::IParallelSuperMng.

Definition at line 550 of file HybridParallelSuperMng.cc.

◆ build()

void Arcane::MessagePassing::HybridParallelSuperMng::build ( )

overridevirtual

Constructs the instance members.

The instance is not usable until this method has been called. This method must be called before initialize().

Warning: This method must only be called once.

Implements Arcane::IParallelSuperMng.

Definition at line 396 of file HybridParallelSuperMng.cc.

References Arcane::IApplication::applicationBuildInfo(), ARCANE_CHECK_POINTER, ARCANE_FATAL, Arcane::arcaneHasThread(), Arcane::arcaneInitializeMPI(), Arcane::arcaneIsCudaAwareMPI(), Arcane::ApplicationInfo::commandLineArguments(), Arcane::ServiceBuilder< InterfaceType >::createReference(), Arcane::RefImpl< InstanceType, RefClassType, ImplTagId >::get(), Arcane::platform::getRealTime(), Arcane::ITraceMng::info(), m_application, m_mpi_comm_rank, m_mpi_comm_size, and Arcane::IBase::traceMng().

Here is the call graph for this function:

◆ commRank()

Int32 Arcane::MessagePassing::HybridParallelSuperMng::commRank ( ) const

inlineoverridevirtual

Returns the process number (between 0 and nbProcess()-1).

Implements Arcane::IParallelSuperMng.

Definition at line 316 of file HybridParallelSuperMng.cc.

References m_mpi_comm_rank.

Referenced by internalCreateWorldParallelMng(), and isMasterIO().

Here is the caller graph for this function:

◆ commSize()

Int32 Arcane::MessagePassing::HybridParallelSuperMng::commSize ( ) const

inlineoverridevirtual

Returns the total number of processes used.

Implements Arcane::IParallelSuperMng.

Definition at line 317 of file HybridParallelSuperMng.cc.

References m_mpi_comm_size.

◆ communicator()

MP::Communicator Arcane::MessagePassing::HybridParallelSuperMng::communicator ( ) const

inlineoverridevirtual

MPI communicator associated with this manager.

See also: IParallelMng::communicator()

Implements Arcane::IParallelSuperMng.

Definition at line 320 of file HybridParallelSuperMng.cc.

◆ getMPICommunicator()

void * Arcane::MessagePassing::HybridParallelSuperMng::getMPICommunicator ( )

inlineoverridevirtual

Address of the MPI communicator associated with this manager.

The communicator is only valid if MPI is used. Otherwise, the returned address is 0. The returned value is of type (MPI_Comm*).

Implements Arcane::IParallelSuperMng.

Definition at line 319 of file HybridParallelSuperMng.cc.

◆ initialize()

void Arcane::MessagePassing::HybridParallelSuperMng::initialize ( )

overridevirtual

Initializes the instance.

The instance is not usable until this method has been called.

Warning: This method must only be called once.

Implements Arcane::IParallelSuperMng.

Definition at line 388 of file HybridParallelSuperMng.cc.

◆ internalCreateWorldParallelMng()

Ref< IParallelMng > Arcane::MessagePassing::HybridParallelSuperMng::internalCreateWorldParallelMng ( Int32 local_rank )

overridevirtual

Creates a parallelism manager for all allocated cores.

This operation is collective.

This method must only be called once during initialization.

local_rank is the local rank of the caller in the list of ranks. In pure MPI mode, this rank is always 0 because there is only one thread. In Thread or Thread/MPI mode, it is the rank of the thread used during creation.

The returned manager remains the property of this instance and must not be destroyed.

For internal use only.

Implements Arcane::IParallelSuperMng.

Definition at line 514 of file HybridParallelSuperMng.cc.

References commRank(), Arcane::ITraceMng::info(), and m_application.

Here is the call graph for this function:

◆ isMasterIO()

bool Arcane::MessagePassing::HybridParallelSuperMng::isMasterIO ( ) const

inlineoverridevirtual

Returns true if the instance is a master I/O manager.

Implements Arcane::IParallelSuperMng.

Definition at line 323 of file HybridParallelSuperMng.cc.

References commRank().

Here is the call graph for this function:

◆ isParallel()

bool Arcane::MessagePassing::HybridParallelSuperMng::isParallel ( ) const

inlineoverridevirtual

Returns true if the execution is parallel.

Implements Arcane::IParallelSuperMng.

Definition at line 315 of file HybridParallelSuperMng.cc.

◆ masterIORank()

Integer Arcane::MessagePassing::HybridParallelSuperMng::masterIORank ( ) const

inlineoverridevirtual

Rank of the instance managing input/output (for which isMasterIO() is true).

In the current implementation, this is always the rank 0 processor.

Implements Arcane::IParallelSuperMng.

Definition at line 324 of file HybridParallelSuperMng.cc.

◆ nbLocalSubDomain()

Integer Arcane::MessagePassing::HybridParallelSuperMng::nbLocalSubDomain ( )

inlineoverridevirtual

Number of subdomains to create locally.

1 if sequential.
1 if pure MPI
n if THREAD or THREAD/MPI

Implements Arcane::IParallelSuperMng.

Definition at line 325 of file HybridParallelSuperMng.cc.

◆ threadMng()

IThreadMng * Arcane::MessagePassing::HybridParallelSuperMng::threadMng ( ) const

inlineoverridevirtual

Thread manager.

Implements Arcane::IParallelSuperMng.

Definition at line 314 of file HybridParallelSuperMng.cc.

◆ traceRank()

Int32 Arcane::MessagePassing::HybridParallelSuperMng::traceRank ( ) const

inlineoverridevirtual

Rank of this instance for traces.

Implements Arcane::IParallelSuperMng.

Definition at line 318 of file HybridParallelSuperMng.cc.

References m_mpi_comm_rank.

◆ tryAbort()

void Arcane::MessagePassing::HybridParallelSuperMng::tryAbort ( )

overridevirtual

Attempts to abort.

This method is called when an exception has been generated and the current execution case must stop. It allows performing cleanup operations on the manager if necessary.

Implements Arcane::IParallelSuperMng.

Definition at line 540 of file HybridParallelSuperMng.cc.

References m_application.

Member Data Documentation

◆ m_application

IApplication* Arcane::MessagePassing::HybridParallelSuperMng::m_application

Main manager.

Definition at line 334 of file HybridParallelSuperMng.cc.

Referenced by application(), build(), internalCreateWorldParallelMng(), and tryAbort().

◆ m_builder_factory

Ref<IParallelMngContainerFactory> Arcane::MessagePassing::HybridParallelSuperMng::m_builder_factory

Definition at line 331 of file HybridParallelSuperMng.cc.

◆ m_communicator

MP::Communicator Arcane::MessagePassing::HybridParallelSuperMng::m_communicator

Definition at line 339 of file HybridParallelSuperMng.cc.

◆ m_container

HybridParallelMngContainer* Arcane::MessagePassing::HybridParallelSuperMng::m_container = nullptr

Definition at line 330 of file HybridParallelSuperMng.cc.

◆ m_datatype_list

MpiDatatypeList* Arcane::MessagePassing::HybridParallelSuperMng::m_datatype_list

Definition at line 344 of file HybridParallelSuperMng.cc.

◆ m_error_handler

MpiErrorHandler Arcane::MessagePassing::HybridParallelSuperMng::m_error_handler

Definition at line 345 of file HybridParallelSuperMng.cc.

◆ m_local_nb_rank

Int32 Arcane::MessagePassing::HybridParallelSuperMng::m_local_nb_rank = A_NULL_RANK

Definition at line 341 of file HybridParallelSuperMng.cc.

◆ m_machine_communicator

MP::Communicator Arcane::MessagePassing::HybridParallelSuperMng::m_machine_communicator

Definition at line 340 of file HybridParallelSuperMng.cc.

◆ m_main_builder

Ref<IParallelMngContainer> Arcane::MessagePassing::HybridParallelSuperMng::m_main_builder

Definition at line 332 of file HybridParallelSuperMng.cc.

◆ m_mpi_adapter

MpiAdapter* Arcane::MessagePassing::HybridParallelSuperMng::m_mpi_adapter

Definition at line 343 of file HybridParallelSuperMng.cc.

◆ m_mpi_comm_rank

Int32 Arcane::MessagePassing::HybridParallelSuperMng::m_mpi_comm_rank = -1

Statistics.

Current processor number

Definition at line 336 of file HybridParallelSuperMng.cc.

Referenced by build(), commRank(), and traceRank().

◆ m_mpi_comm_size

Int32 Arcane::MessagePassing::HybridParallelSuperMng::m_mpi_comm_size = -1

Number of processors.

Definition at line 337 of file HybridParallelSuperMng.cc.

Referenced by build(), and commSize().

◆ m_mpi_communicator

MPI_Comm Arcane::MessagePassing::HybridParallelSuperMng::m_mpi_communicator

Definition at line 338 of file HybridParallelSuperMng.cc.

◆ m_mpi_lock

MpiLock* Arcane::MessagePassing::HybridParallelSuperMng::m_mpi_lock

Definition at line 342 of file HybridParallelSuperMng.cc.

◆ m_stat

Parallel::IStat* Arcane::MessagePassing::HybridParallelSuperMng::m_stat

Definition at line 335 of file HybridParallelSuperMng.cc.

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

HybridParallelSuperMng.cc

Public Member Functions

Public Attributes

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ HybridParallelSuperMng()

◆ ~HybridParallelSuperMng()

Member Function Documentation

◆ application()

◆ barrier()

◆ build()

◆ commRank()

◆ commSize()

◆ communicator()

◆ getMPICommunicator()

◆ initialize()

◆ internalCreateWorldParallelMng()

◆ isMasterIO()

◆ isParallel()

◆ masterIORank()

◆ nbLocalSubDomain()

◆ threadMng()

◆ traceRank()

◆ tryAbort()

Member Data Documentation

◆ m_application

◆ m_builder_factory

◆ m_communicator

◆ m_container

◆ m_datatype_list

◆ m_error_handler

◆ m_local_nb_rank

◆ m_machine_communicator

◆ m_main_builder

◆ m_mpi_adapter

◆ m_mpi_comm_rank

◆ m_mpi_comm_size

◆ m_mpi_communicator

◆ m_mpi_lock

◆ m_stat