MPI Send/Receive/Scatter for DREAM

Collaboration diagram for MPI Send/Receive/Scatter for DREAM:

Functions
template<class Likelihood >
int	TasDREAM::MPILikelihoodSend (Likelihood const &likely, int destination, int tag, MPI_Comm comm, int outputs_begin=0, int outputs_end=-1)
	Send a likelihood to another process in the MPI comm. More...
template<class Likelihood >
int	TasDREAM::MPILikelihoodRecv (Likelihood &likely, int source, int tag, MPI_Comm comm, MPI_Status *status=MPI_STATUS_IGNORE)
	Receive a likelihood from another process in the MPI comm. More...
template<class Likelihood >
int	TasDREAM::MPILikelihoodScatter (Likelihood const &source, Likelihood &destination, int root, int tag, MPI_Comm comm)
	Split the likelihood across the comm where each rank receives an equal portion of the total outputs. More...

Detailed Description

Methods to send/receive DREAM likelihood objects. The syntax mimics the raw MPI_Send and MPI_Recv calls, and the templates require Tasmanian_ENABLE_MPI=ON.

Function Documentation

MPILikelihoodSend()

template<class Likelihood >

int TasDREAM::MPILikelihoodSend	(	Likelihood const &	likely,
		int	destination,
		int	tag,
		MPI_Comm	comm,
		int	outputs_begin = 0,
		int	outputs_end = -1
	)

Send a likelihood to another process in the MPI comm.

Works with both isotropic and anisotropic Gaussian likelihood object implemented in Tasmanian. The usage is very similar to MPI_Send() and TasGrid::MPIGridSend(). Optionally only some of the outputs can be send over.

Template Parameters

Likelihood

is a Tasmanian likelihood class, currently LikeliTasDREAM::LikelihoodGaussIsotropic and LikeliTasDREAM::LikelihoodGaussAnisotropic.

Parameters

likely	is the likelihood to send.
destination	is the rank of the recipient MPI process.
tag	is the tag to use for the MPI message.
comm	is the MPI comm where the source and destination reside.
outputs_begin	same as in LikelihoodGaussIsotropic::write().
outputs_end	same as in LikelihoodGaussIsotropic::write().

Returns

the error code of the MPI_Send() command.

Note: this call must be mirrored by TasDREAM::MPILikelihoodRecv() on the destination process.

MPILikelihoodRecv()

template<class Likelihood >

int TasDREAM::MPILikelihoodRecv	(	Likelihood &	likely,
		int	source,
		int	tag,
		MPI_Comm	comm,
		MPI_Status *	status = MPI_STATUS_IGNORE
	)

Receive a likelihood from another process in the MPI comm.

Works with both isotropic and anisotropic Gaussian likelihood object implemented in Tasmanian. The usage is very similar to MPI_Recv() and TasGrid::MPIGridRecv().

Template Parameters

Likelihood

is a Tasmanian likelihood class, currently LikeliTasDREAM::LikelihoodGaussIsotropic and LikeliTasDREAM::LikelihoodGaussAnisotropic.

Parameters

likely	is the output likelihood, it will be overwritten with the one send.
source	is the rank of the process in the MPI comm that issued the send command.
tag	is the tag used in the MPI send command.
comm	is the MPI comm where the source and destination reside.
status	is the status of the MPI_Recv() command.

Returns

the error code of the MPI_Recv() command.

Note: see TasDREAM::MPILikelihoodSend().

MPILikelihoodScatter()

template<class Likelihood >

int TasDREAM::MPILikelihoodScatter	(	Likelihood const &	source,
		Likelihood &	destination,
		int	root,
		int	tag,
		MPI_Comm	comm
	)

Split the likelihood across the comm where each rank receives an equal portion of the total outputs.

Splits both the data and the variance across a comm.

Note: this does not use MPI_Scatter(), instead it makes multiple calls to MPILikelihoodSend() and MPILikelihoodRecv().

Template Parameters

Likelihood

is a Tasmanian likelihood class, currently LikeliTasDREAM::LikelihoodGaussIsotropic and LikeliTasDREAM::LikelihoodGaussAnisotropic.

Parameters

source	likelihood located on the root rank is to be distributed across, for all other ranks the source will not be referenced.
destination	is the likelihood where the local portion of the scatter will be stored, the existing likelihood will be overwritten. If the source outputs are less than the number of comm ranks, then some of the destination likelihoods will be empty.
root	is the rank that will hold the source likelihood.
tag	same as in TasGrid::MPILikelihoodSend().
comm	is the MPI comm of all process that need to share a portion of the grid.

Note: see TasGrid::MPIGridScatterOutputs() for the way the outputs will be distributed.